CN105260749A - Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM - Google Patents

Abstract

The invention provides a real-time target detection method based on the oriented gradient two-value mode (ORBP) and soft cascade SVM, and aims at solving the problems that the target detection in the prior art is low in instantaneity and robustness. The method comprises the steps that 1) characteristic of the oriented gradient two-value mode is described; 2) a soft cascade classifier SVM is established; 3) characteristic training is carried out on the soft cascade classifier SVM; and 4) a target window is tracked and updated. ORBP has the advantages that rotation, scale, translation and brightness are not changed, the soft cascade SVM improves the target detection robustness in complex scenes, and tracking of the target window improves the instantaneity of target detection. The method provided by the invention can be applied to man-machine interaction and intelligent traffic monitoring fields, and the target detection performance is excellent.

Description

Real-time Object Detection Method Based on Oriented Gradient Binary Mode and Soft Cascaded SVM

technical field

The present invention relates to the technical field of digital image processing, relates to target detection and tracking methods in computer vision, can be applied to the fields of human-computer interaction and intelligent transportation, and specifically relates to a real-time Object detection method.

Background technique

Target detection is to automatically analyze the image to detect the target of interest through computer information processing technology. As an important topic of image understanding, target detection is widely used in military and civilian scenarios. In the real scene, due to the presence of other interfering moving objects in the background of the scene, changes in the external lighting environment, and various and rapid changes in the shape of the target, it brings many difficulties to target detection. How to achieve efficient and stable target detection has important practical research significance.

Zhang Tianyu proposed a multi-scale target detection method in the patent "Space-Time Multi-Scale Moving Target Detection Method", which divides the image into blocks and uses the optimal difference interval in the moving area to achieve target detection and tracking. This method is robust in complex scenes. The significance is low, and the criterion of significant difference is difficult to adapt to multiple scenarios. Zdenek Kalal, Krystian Mikolajczyk and others proposed a single target detection and tracking method in the video in "Tracking-Learning-Detection", which uses the information difference between frames to combine detection and tracking to achieve online learning of target samples. The proposed median optical flow method requires target initialization, and it is difficult to ensure synchronization with the detector when the tracking correction is fixed. Yang Yanshuang and Pu Baoming proposed the adaptive threshold SUSAN method to detect the vehicle target boundary in "Moving Vehicle Detection Based on Improved SUSAN Algorithm", using the combination of histogram transformation and Hough transform to extract the target connected domain, and realize the detection of the vehicle target and the background. Separation, the real-time performance of this method is poor and it will be difficult to effectively complete the target segmentation in complex scenes with adaptive threshold.

Contents of the invention

Aiming at the deficiencies of the above-mentioned prior art, the present invention proposes a real-time target detection method based on directional gradient binary mode and soft cascaded SVM in order to solve the problems of low robustness and poor real-time performance of existing target detection methods in complex scenes , with excellent target detection performance and easy engineering implementation.

The above objects of the invention are achieved by the technical features of the independent claims, which the dependent claims develop in an alternative or advantageous manner.

In order to achieve the above object, the present invention provides a real-time target detection method based on directional gradient binary mode and soft cascaded SVM. Target feature description, during feature training, use the random position of the detection image to generate positive and negative samples, and finally use shi-Tomasi corner point detection to extract feature points to complete the target tracking update.

In some embodiments, the real-time target detection method includes the following steps:

(1) ORBP feature extraction. Perform preprocessing operations on image source samples, and use Sobel edges and local directional gradients to generate ORBP features.

(2) Construction of soft cascade classifier SVM. Use the similarity of cross-correlation features to determine whether the feature selection of the sample is valid, calculate the responses of all samples according to h _k (x), and find the threshold corresponding to the positive sample boundary classification. The threshold and features corresponding to this level will be added to the k+1 level Calculate the response h _k+1 (x); then send the window set to be detected to the soft cascade classifier in turn, and judge whether it belongs to the target by judging the response of the current window.

(3) Training of soft cascaded classifiers. Generate positive samples and negative samples for the calibrated positive sample target image, and perform ORBP feature description on the sample; then train the initial classifier h ₀ (x) through SVM, perform target detection and verification on the sample image according to the initial classifier, and convert the negative sample Re-update to the next level of SVM cascade classifier training until the final cascade classifier training is completed.

(4) Target window tracking update. According to the classifier trained by the soft cascaded SVM, the target window detection is performed on the image sequence to be detected, and the feature points of the target window are extracted by using the shi-Tomasi corner detection method, and whether the current feature point is the best tracking is determined according to the Median-Flow tracker point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to judge the target, and finally output the target detection window.

Wherein, the ORBP feature generation described in step (1) comprises the following steps:

(1) Preprocess the contrast transformation of the image source sample to eliminate the influence of ambient light. The preprocessing operation includes Gaussian smoothing filter and Gamma normalization transformation.

(2) Divide the gradient direction into K parts, calculate the gradient amplitudes in each gradient direction of the Sobel vertical edge, put the gradient modulus lengths corresponding to the K direction intervals into M sub-block matrices, and generate corresponding edges Oriented Gradient Map.

(3) For each edge direction gradient map, divide the edges in the horizontal and vertical directions, and count the cumulative responses in the horizontal and vertical directions respectively. For the horizontal direction, the cumulative response on the upper side is m ₁ , and the cumulative response on the lower side is m ₂ ; for the vertical direction, the cumulative response on the left side is m ₃ , and the cumulative response on the right side is m ₄ , as shown in Figure 2.

(4) Generate the ORBP feature based on the comparison of the cumulative response in the horizontal direction up and down and the cumulative response in the vertical direction. As shown in Figure 2, the generated ORBP feature contains 4 binary forms, and the direction gradient histogram will be converted when describing the feature It is one of the corresponding ORBP forms.

Wherein, the detailed process of the cross-correlation feature similarity determination method described in step (2) is as follows:

Perform ORBP feature extraction on image source samples, randomly select one-dimensional features as the initial feature f ₀ , add other one-dimensional features as the second feature f ₁ , calculate the normalized cross-correlation coefficient η between features f ₀ and f ₁ , and normalize The calculation method of the normalized cross-correlation coefficient η is as follows:

$\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; cov</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\sqrt{\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}}$

Where f _i represents the i-th dimension feature vector, cov(f _i , f _j ) represents the covariance of feature vector f _i and f _j , and var(f _i ) represents the variance of feature vector f _i . According to the calculated cross-correlation coefficient η, if η<0.6, it is determined that the currently added feature is valid, otherwise it is determined that the current feature is invalid, and any dimension feature needs to be selected again. If you continue to add any dimension feature f _i+1 , you need to judge whether the cross-correlation coefficient with the current feature vector set {f ₀ , f ₁ ...f _i } meets the condition.

Further, in order to reduce the time complexity of feature description, in the method (1) ORBP feature description, only the vertical edge is extracted from the Sobel edge.

Further, in order to improve the completeness of sample selection, the selection of the position window in step (3) during the positive and negative sample generation process of this method needs to satisfy: select 10 bounding boxes closest to it from the neighborhood of the target, each scale Select up to 4 position windows as positive and negative samples.

Beneficial effects: the present invention proposes a real-time target detection method based on the directional gradient binary mode and soft cascaded SVM to solve the problems of low robustness and poor real-time performance of target detection, and uses the directional gradient binary mode as a feature descriptor to improve feature description Robustness to complex scene backgrounds and illumination changes; at the same time, adaptive feature selection is used to construct soft cascaded SVM multi-level classification thresholds, and random positive and negative samples are used to train classifiers and track target windows, and soft cascades reduce target window screening , window tracking improves the stability and real-time performance of multi-frame sequence detection. Compared with other similar target detection algorithms, the method proposed by the invention has strong robustness, good real-time performance, and excellent target detection performance.

It should be understood that all combinations of the foregoing concepts, as well as additional concepts described in more detail below, may be considered part of the inventive subject matter of the present disclosure, provided such concepts are not mutually inconsistent. Additionally, all combinations of claimed subject matter are considered a part of the inventive subject matter of this disclosure.

The foregoing and other aspects, embodiments and features of the present teachings can be more fully understood from the following description when taken in conjunction with the accompanying drawings. Other additional aspects of the invention, such as the features and/or advantages of the exemplary embodiments, will be apparent from the description below, or learned by practice of specific embodiments in accordance with the teachings of the invention.

Description of drawings

The figures are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures may be represented by a like reference numeral. For purposes of clarity, not every component may be labeled in every drawing. Embodiments of the various aspects of the invention will now be described by way of example with reference to the accompanying drawings, in which:

FIG. 1 is a flow chart of a real-time object detection method based on directional gradient binary mode and soft cascaded SVM according to some embodiments of the present invention.

Fig. 2 is a schematic diagram of ORBP feature generation according to some embodiments of the present invention.

Fig. 3 is a schematic diagram of multi-target detection according to some embodiments of the present invention.

Fig. 4 is a schematic diagram of object detection in complex scenes according to some embodiments of the present invention.

detailed description

In order to better understand the technical content of the present invention, specific embodiments are given together with the attached drawings for description as follows.

Aspects of the invention are described in this disclosure with reference to the accompanying drawings, which show a number of illustrated embodiments. Embodiments of the present disclosure are not necessarily intended to include all aspects of the invention. It should be appreciated that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of numerous ways, since the concepts and embodiments disclosed herein are not limited to any implementation. In addition, some aspects of the present disclosure may be used alone or in any suitable combination with other aspects of the present disclosure.

As shown in FIG. 1, according to an embodiment of the present invention, the real-time target detection method based on directional gradient binary mode and soft cascade SVM includes the following steps:

1. Oriented gradient binary pattern (ORBP) feature description

1.1 Preprocessing the image source f(x,y) sample contrast transformation, the preprocessing operation includes Gaussian smoothing filtering and normalization processing, here the image normalization operation uses Gamma normalization transformation:

f(x,y)=ln(f(x,y)+1)

1.2 Divide the gradient direction [-π/2, π/2] into 9 intervals, calculate the gradient amplitude |▽f| in each gradient direction under the Sobel vertical edge, and generate the corresponding edge direction gradient map. The gradient magnitude |▽f| and the direction angle θ are calculated as follows:

$<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&dtri;</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; =</span>\sqrt{{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

θ=arctan(G _y /G _x )

Among them, G _x and G _y are the gradients of the image f(x, y) along the x and y directions, respectively.

1.3 Divide the image into 100 square blocks, each unit block is composed of 6×6 square cells, respectively divide the edge direction gradient of each square block in the horizontal and vertical directions, and calculate the cumulative responses in the horizontal and vertical directions. For the horizontal direction, the cumulative response on the upper side is m ₁ , and the cumulative response on the lower side is m ₂ ; for the vertical direction, the cumulative response on the left side is m ₃ , and the cumulative response on the right side is m ₄ , as shown in Figure 2.

1.4 Generate local binary features based on the comparison of the cumulative responses in the horizontal direction up and down and the cumulative responses in the vertical direction. As shown in Figure 2, the generated ORBP features contain 4 binary forms, which will be converted into the corresponding ORBP forms when describing the features kind of.

2. Construction of soft cascade classifier SVM

2.1 Judging the similarity of cross-correlation features, according to step 1, obtain the binary pattern features of the direction gradient, randomly select one-dimensional features as the initial feature f ₀ , add other one-dimensional features as the second feature f ₁ , and calculate the features f ₀ and f ₁ The normalized cross-correlation coefficient η, the calculation method of the normalized cross-correlation coefficient η is as follows:

$\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; cov</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\sqrt{\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}}$

Where f _i represents the i-th dimension feature vector, cov(f _i , f _j ) represents the covariance of feature vector f _i and f _j , and var(f _i ) represents the variance of feature vector f _i . According to the calculated cross-correlation coefficient η, if η<0.6, it is determined that the currently added feature is valid, otherwise it is determined that the current feature is invalid, and any dimension feature needs to be selected again. If you continue to add any dimension feature f _i+1 , you need to judge whether the cross-correlation coefficient with the current feature vector set {f ₀ , f ₁ ...f _i } meets the condition.

2.2 Generation of cascade thresholds, constructing a soft cascade classifier h _k (x) of n-dimensional linear SVM:

${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

Among them, w _i is the support vector of classification decision plane i, and x _i is the corresponding i-dimensional feature. The cascade classifier judges whether it is valid according to the nearest neighbor feature at each level of feature selection, and then calculates the response of all samples according to h _k (x), and finds the threshold corresponding to the positive sample boundary classification. The corresponding threshold and feature of this level will be Will be added to level k+1 to calculate the response h _k+1 (x).

2.3 Determination of cascade classification. The window set to be detected is sent to the soft cascade classifier in turn, and the window to be detected is screened according to the threshold and characteristics obtained in the previous step. When the response of the current window is less than the decision threshold, it will be considered as a non-target, and then the remaining windows will be processed in the next step. One-level cascade classification decision. If the response of the current window is higher than the decision threshold of the current classifier, the current window is considered as the target.

3. Soft cascade classifier feature training

3.1 Positive sample generation, scan the calibrated positive sample target image with windows of different scales, select 10 bounding boxes closest to it from the vicinity of the target neighborhood, and judge whether the currently selected positive sample is positive according to the size of the window intersection area ratio. Meet the criteria. It should be noted that at each scale, at most 4 position windows are selected as positive samples to ensure the completeness of the training samples.

3.2 Negative sample generation. While positive samples are selected, the position of the positive sample window is translated up, down, left, and right. When the window intersection area is lower than a certain threshold, the sample generated by the current window is considered to be a negative sample. Each scale also generates up to 4 position windows as Negative samples.

3.3 Initial feature training, the training sample set is obtained according to the positive and negative samples, 200 positive and negative samples are randomly selected, and the lower samples of different scales are uniformly normalized to 60×60, and the direction gradient binary mode feature description is performed on the samples. An initial classifier h ₀ (x) is trained.

3.4 Cascading feature training, perform object detection on all images according to the initial soft cascade, and add false detection windows to the next level of classifier feature training according to the position of the positive sample window. Train the next level classification h _k (x), k∈[2,+∞), until the update termination condition is satisfied.

4. Target window tracking update

4.1 Cascade target window detection. According to the threshold and the corresponding feature vector of the trained soft cascade SVM classifier, the image sequence is judged step by step, and the final cascade decision judgment output window is the target window of the current frame.

4.2 Extract target window feature points. In the above step, the output window of the cascade classifier is obtained, and the feature points of the target window are extracted by using the shi-Tomasi corner detection method. It should be noted that for the real-time and accuracy of follow-up tracking, the worst quality assurance of corner detection is excellent, and the number of corner detections in the control window is not more than 20.

4.3 Target window tracking point screening, according to the tracking point of the current nth frame window, use the Median-Flow tracker to track forward to the n+1th frame, and then reversely track to the nth frame, and calculate the change of the number of corner points in the window and the backtracking point The Euclidean distance from the original corner point, set the judgment threshold to filter the best tracking point.

4.4 Target window update strategy, use the best tracking point to calculate the predicted position of the target window in the next frame, and use the initial classifier of the cascade classifier to determine the target. If the current window is judged as the target, the current tracking is considered valid; otherwise, the sliding window search needs to be performed according to the cascade classifier, and the target detection is performed again.

The real-time target detection method based on the directional gradient binary mode and the soft cascaded SVM proposed by the present invention will be further tested in detail in combination with specific scenarios below. Implement this example method on hardware platform Interi5+4GDDR3RAM, software platform OpenCV/C ++, scene multi-target detection is tested in the accompanying drawing 3, present embodiment reaches 96% to the target detection rate, and the single frame running time is 27ms; Accompanying drawing 4 Vehicle target detection in medium and complex scenes is tested. In this embodiment, the target detection rate reaches 95.3%, and the running time of a single frame is 25ms.

From the above technical solutions, it can be known that the real-time target detection method based on the directional gradient binary mode and the soft cascaded SVM provided by the present invention is based on the soft cascaded support vector machine SVM, and adopts the feature based on the directional gradient binary mode. Use target feature description to improve the robustness of feature description to complex scene background and illumination changes; during feature training, use the random position of the detection image to generate positive and negative samples, and finally use shi-Tomasi corner point detection to extract feature points to complete target tracking update , soft cascading reduces target window screening, and window tracking improves the stability and real-time performance of multi-frame sequence detection. The method proposed by the invention has strong robustness, high real-time performance and excellent target detection performance.

According to the present disclosure, a real-time target detection device based on a binary mode of directional gradient and soft cascaded SVM is also proposed, the device includes:

A first module for ORBP feature extraction, the first module is configured to perform preprocessing operations on image source samples, using Sobel edges and local direction gradients to generate ORBP features;

For constructing the second module of the soft cascade classifier SVM, the second module is set to use the cross-correlation feature similarity to determine whether the sample feature selection is valid, calculate the responses of all samples according to h _k (x), and find the positive sample The threshold corresponding to the boundary classification, the threshold and features corresponding to this level will be added to the k+1 level to calculate the response h _k+1 (x); Response to determine whether it belongs to the target. ;

The third module for training the soft cascade classifier, the third module is set to generate positive samples and negative samples for the calibration positive sample target image, and perform ORBP feature description on the samples; then train the initial classifier by SVM h ₀ (x), perform target detection verification on the sample image according to the initial classifier, and update the negative sample to the next level of SVM cascade classifier training until the final cascade classifier training is completed;

The fourth module for target window tracking update, the fourth module is set to detect the target window of the image sequence to be detected according to the classifier trained by the soft cascade SVM, and extract the target window by using the shi-Tomasi corner detection method Feature points, according to the Median-Flow tracker to determine whether the current feature point is the best tracking point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to determine the target, and finally Output object detection window.

It should be understood that the functions, functions and effects of the first module, the second module, the third module and the fourth module proposed in this embodiment have been described above in the real-time target detection method based on the direction gradient binary mode and soft cascaded SVM It is explained in the description of , and its implementation is illustrated in the aforementioned embodiments of the real-time target detection method, and will not be repeated here.

According to the aforementioned embodiments of the present invention, such as a real-time target detection method based on a binary mode of directional gradient and soft cascaded SVM and a real-time target detection device based on a binary mode of directional gradient and soft cascaded SVM, the present invention also proposes a method using A computer system for realizing real-time target detection based on directional gradient binary mode and soft cascade SVM, the computer system includes:

memory;

one or more processors;

One or more modules, the one or more modules are stored in the memory and configured to be executed by the one or more processors, the one or more modules include the following processing modules:

A first module for ORBP feature extraction, the first module is configured to perform preprocessing operations on image source samples, using Sobel edges and local direction gradients to generate ORBP features;

For constructing the second module of the soft cascade classifier SVM, the second module is set to use the cross-correlation feature similarity to determine whether the sample feature selection is valid, calculate the responses of all samples according to h _k (x), and find the positive sample The threshold corresponding to the boundary classification, the threshold and features corresponding to this level will be added to the k+1 level to calculate the response h _k+1 (x); Response to determine whether it belongs to the target. ;

The third module for training the soft cascade classifier, the third module is set to generate positive samples and negative samples for the calibration positive sample target image, and perform ORBP feature description on the samples; then train the initial classifier by SVM h ₀ (x), perform target detection verification on the sample image according to the initial classifier, and update the negative sample to the next level of SVM cascade classifier training until the final cascade classifier training is completed;

The fourth module for target window tracking update, the fourth module is set to detect the target window of the image sequence to be detected according to the classifier trained by the soft cascade SVM, and extract the target window by using the shi-Tomasi corner detection method Feature points, according to the Median-Flow tracker to determine whether the current feature point is the best tracking point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to determine the target, and finally Output object detection window.

Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the claims.

Claims (7)

1. a real-time target detection method based on directional gradient binary mode and soft cascade SVM, characterized in that, comprising the following steps: (1) ORBP feature extraction: perform contrast transformation preprocessing on the image source sample, divide the gradient direction into K parts, and calculate the gradient map of each direction block under the edge of the image Sobel respectively; then according to the horizontal and vertical accumulation of the direction block gradient map In response, generate ORBP features; (2) Construction of soft cascade classifier SVM: According to the ORBP feature of the image source sample, calculate the response of all samples, find the threshold and feature vector corresponding to the positive sample boundary classification, and then send the windows to be detected to the soft cascade classification in turn The device judges whether it belongs to the target through the current window response size; (3) Soft cascade classifier feature training: generate positive and negative samples for the calibrated positive sample target image, randomly select N positive and negative samples each, describe the ORBP features of the samples, and then use the constructed soft cascade SVM classifier to complete Training on sample features; (4) Target window tracking update: according to the classifier trained by the soft cascade SVM, the target window detection is performed on the image sequence, and the feature points of the target window are extracted by using the shi-Tomasi corner detection method, and the current state is determined according to the Median-Flow tracker. Whether the feature point is the best tracking point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to judge the target, and finally output the target detection window. 2. the real-time target detection method based on direction gradient binary mode and soft cascade SVM as claimed in claim 1, is characterized in that, the concrete method of ORBP feature extraction in described step (1) comprises the following steps: (1) Contrast transformation preprocessing operation is performed on the image source sample, and the preprocessing operation includes Gaussian smoothing filtering and contrast normalization processing; (2) Divide the gradient direction into K parts, respectively calculate the gradient amplitude under each gradient direction under the Sobel edge, put the gradient modulus length of the K direction range into the corresponding M sub-block matrix, and generate the corresponding edge direction block Gradient map; (3) Carry out horizontal and vertical direction edge divisions for each edge direction gradient map, and count the horizontal and vertical direction cumulative responses respectively; (4) Generate ORBP features based on the comparison of the cumulative response in the horizontal direction up and down and the cumulative response in the vertical direction. 3. as claimed in claim 1, based on the real-time target detection method of direction gradient binary mode and soft cascade SVM, it is characterized in that, the construction of soft cascade SVM in the described step (2), concrete realization comprises: (1) Calculating cross-correlation feature similarity judgment: extracting ORBP features for image source samples, using cross-correlation to calculate the similarity between features, and judging whether the sample feature selection is valid according to the normalized cross-correlation coefficient; (2) Generation of cascade thresholds: Construct a soft cascade classifier h _k (x) of n-dimensional linear SVM: ${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$ Among them, w _i is the support vector of classification decision plane _i , and x _i is the corresponding i-dimensional feature; the cascade classifier judges whether it is valid according to the cross-correlation feature when selecting features at each level, and then calculates all For the response of the sample, find the threshold corresponding to the positive sample boundary classification, and the threshold and features corresponding to this level will be added to the k+1 level to calculate the response h _k+1 (x); (3) Judgment of cascade classification: Send all windows to be detected to the soft cascade classifier in turn, and use the threshold and features obtained by cascade to screen the detection windows. When the response of the current window is less than the decision threshold, it will be considered true or false. target, and then the remaining windows are judged by the next cascade classifier: if the response of the current window is higher than the decision threshold of the current classifier, the current window is considered as the target. 4. as claimed in claim 1, based on the real-time target detection method of directional gradient binary mode and soft cascade SVM, it is characterized in that, in the described step (3), generating positive and negative samples specifically includes: 1) Positive sample generation: Scan the calibrated positive sample target image with windows of different scales, select 10 bounding boxes closest to it from the vicinity of the target neighborhood, and determine the currently selected positive sample according to the size of the window intersection area ratio whether it meets the conditions; 2) Negative sample generation: While positive samples are selected, the position of the positive sample window is translated up, down, left, and right. When the window intersection area is lower than a certain threshold, the sample generated by the current window is considered to be a negative sample, and each scale also generates up to 4 position windows. as a negative sample. 5. as claimed in claim 3, based on the real-time target detection method of directional gradient binary mode and soft cascade SVM, it is characterized in that, the specific method for determining the similarity of said cross-correlation feature is: Perform ORBP feature extraction on image source samples, randomly select one-dimensional features as the initial feature f ₀ , add other one-dimensional features as the second feature f ₁ , calculate the normalized cross-correlation coefficient η between features f ₀ and f ₁ , and normalize The calculation method of the normalized cross-correlation coefficient η is as follows: $\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; cov</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\sqrt{\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; var</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}}$ Among them, f _i represents the i-th dimension feature vector, cov(f _i , f _j ) represents the covariance of feature vector f _i and f _j , var(f _i ) represents the variance of feature vector f _i ; according to the calculated cross-correlation coefficient η, if η<η ₀ , it is determined that the currently added feature is valid, otherwise it is determined that the current feature is invalid, and any dimension feature needs to be _reselected ; Whether the cross-correlation coefficient of {f ₀ ,f ₁ ...f _i } satisfies the condition. 6. A real-time target detection device based on directional gradient binary mode and soft cascade SVM, characterized in that the device comprises: A first module for ORBP feature extraction, the first module is configured to perform preprocessing operations on image source samples, using Sobel edges and local direction gradients to generate ORBP features; For constructing the second module of the soft cascade classifier SVM, the second module is set to use the cross-correlation feature similarity to determine whether the sample feature selection is valid, calculate the responses of all samples according to h _k (x), and find the positive sample The threshold corresponding to the boundary classification, the threshold and features corresponding to this level will be added to the k+1 level to calculate the response h _k+1 (x); Response to determine whether it belongs to the target; The third module for training the soft cascade classifier, the third module is set to generate positive samples and negative samples for the calibration positive sample target image, and perform ORBP feature description on the samples; then train the initial classifier by SVM h ₀ (x), perform target detection verification on the sample image according to the initial classifier, and update the negative sample to the next level of SVM cascade classifier training until the final cascade classifier training is completed; The fourth module for target window tracking update, the fourth module is set to detect the target window of the image sequence to be detected according to the classifier trained by the soft cascade SVM, and extract the target window by using the shi-Tomasi corner detection method Feature points, according to the Median-Flow tracker to determine whether the current feature point is the best tracking point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to determine the target, and finally Output object detection window. 7. A computer system for realizing the real-time target detection based on directional gradient binary mode and soft cascade SVM, characterized in that, the computer system includes: memory; one or more processors; One or more modules, the one or more modules are stored in the memory and configured to be executed by the one or more processors, the one or more modules include the following processing modules: A first module for ORBP feature extraction, the first module is configured to perform preprocessing operations on image source samples, using Sobel edges and local direction gradients to generate ORBP features; For constructing the second module of the soft cascade classifier SVM, the second module is configured to use the cross-correlation feature similarity to determine whether the sample feature selection is valid, and calculate the values of all samples according to the soft cascade classifier h _k (x) Response, find the threshold corresponding to the positive sample boundary classification, the threshold and features corresponding to this level will be added to the k+1 level to calculate the response h _k+1 (x); then the window set to be detected is sent to the soft cascade classifier in turn , by judging the current window response to judge whether it belongs to the target; The third module for training the soft cascade classifier, the third module is set to generate positive samples and negative samples for the calibration positive sample target image, and perform ORBP feature description on the samples; then train the initial classifier by SVM h ₀ (x), perform target detection verification on the sample image according to the initial classifier, and update the negative sample to the next level of SVM cascade classifier training until the final cascade classifier training is completed; The fourth module for target window tracking update, the fourth module is set to detect the target window of the image sequence to be detected according to the classifier trained by the soft cascade SVM, and extract the target window by using the shi-Tomasi corner detection method Feature points, according to the Median-Flow tracker to determine whether the current feature point is the best tracking point; then calculate the predicted position of the target window in the next frame through the best tracking point, use the initial classifier of the cascade classifier to determine the target, and finally Output object detection window.