A target detection and automatic tracking method based on intelligent selection strategy

Abstract

The present invention discloses a target detection and automatic tracking algorithm based on an intelligent selection strategy, which belongs to the field of computer vision. The method is composed of a multi-scale detection network, an intelligent selection strategy, a template matching network and a prediction module. First, a multi-scale detection network is used to detect the target in real time, and then the detection candidate frame is selected according to the intelligent selection strategy. Different influencing factors can be set according to different needs, and it has high flexibility and scalability. Finally, the target of interest is tracked in real time through the template matching network. At the same time, in order to solve the problem that the target is blocked or temporarily disappears, the present invention proposes a prediction module to predict the target. Finally, through quantitative evaluation tests, it is shown that under complex background environments, the algorithm proposed by the present invention realizes the functions of real-time detection, intelligent selection and automatic tracking of the target, which has a good promoting effect and reference value on the development of the field of computer vision.

Description

Technical Field

[0001] The present invention relates to a field, and in particular to a target detection and automatic tracking algorithm based on an intelligent selection strategy. Background Art

[0002] Visual target detection and tracking is a research hotspot in the field of computer vision, and has high application value in the fields of PTZ tracking, unmanned aerial vehicles, intelligent video surveillance, and intelligent transportation systems. Since the traditional single-target tracking algorithm cannot detect and intelligently select the target of interest, and the multi-target tracking algorithm also has shortcomings such as missed detection and false detection. Therefore, it is necessary to design a target detection and automatic tracking algorithm based on an intelligent selection strategy to achieve real-time detection, intelligent selection, and automatic tracking of the target, which has a good promoting effect and reference value for the development of the computer vision field. Summary of the invention

[0003] The purpose of the present invention is to provide a target detection and automatic tracking algorithm based on an intelligent selection strategy, which solves the technical problems that the traditional single target tracking algorithm cannot detect and intelligently select the target of interest, and the multi-target tracking algorithm still has shortcomings such as missed detection and false detection. Tests and experiments show that compared with the existing target tracking method, the algorithm proposed in the present invention realizes the functions of real-time detection, intelligent selection and automatic tracking of the target, and can adapt to various complex applications.

[0004] An object detection and automatic tracking algorithm based on an intelligent selection strategy, the algorithm comprises the following steps:

[0005] Step 1: Build a target detection and tracking network model, which consists of a multi-scale detection network, an intelligent selection strategy, a template matching network, and a prediction module;

[0006] Step 2: Train the target detection and tracking network model. Train the multi-scale detection network and template matching network on the COCO dataset and ImageNet dataset respectively.

[0007] Step 3: Test the target detection and tracking network model. Use the trained target detection and tracking network model to perform target detection and tracking tests to achieve intelligent selection and automatic real-time tracking of targets.

[0008] Furthermore, in step 1, the multi-scale detection network is composed of a darknet-53 network and three prediction layers, which are used to extract features of the input video image and predict and locate the target. After the multi-scale detection network completes the prediction of the input video image, the intelligent selection strategy selects the target. The principle of selecting the target by the intelligent selection strategy is composed of four factors, which are: target category, target area, target detection accuracy and target movement speed. After the intelligent selection strategy frames the target of interest, the target of interest is input into the template matching network and tracked in real time. The template matching network is composed of two subnets, which share the same parameters, and the functions of the two subnets are to extract features of the target image and the search image respectively. In a complex background, when the target is deformed or blocked, the prediction module predicts the state and position of the target.

[0009] Furthermore, the darknet-53 network has a total of 53 convolutional layers, which are composed of 1x1 and 3x3 convolutional layers respectively. The performance of Darknet-53 is reflected in the difference in the basic units of the network. It adopts ResNet as the skip layer connection method, and its performance is completely better than resnet-152 and resnet-101. Multi-scale detection enhances the detection of objects of different sizes and blocked objects through the prediction method of three different scale prediction layers. At the same time, darknet-53 introduces the skip layer connection method to enhance convergence, and adopts the random multi-scale training method to enhance robustness.

[0010] Furthermore, the selection principle of the intelligent selection strategy is as follows:

[0011] f(x)=max[ω1c i +ω2s i +ω3a i +ω4v i ] (1)

[0012] Among them, c i is the target category, s i is the target area, a i is the target detection accuracy, v i The above four factors are assigned four corresponding parameters, namely ω1, ω2, ω3 and ω4, which can freely set their own parameter values ​​according to specific requirements. It has high flexibility and scalability. The intelligent selection strategy calculates the value of each target in the video image, and the target with the highest score is the target of interest selected by this strategy.

[0013] Furthermore, the template matching network is composed of subnets, which include a target subnet and a search subnet, which perform feature extraction on the target image and the search image respectively and share the same weights and biases, wherein both the target subnet and the search subnet adopt an internal cropping unit based on a bottleneck residual block, and the internal cropping unit crops out features affected by zero padding inside the block to prevent the convolution filter from learning position deviations, wherein the residual unit is a key module of the template matching network, which is composed of 3 stacked convolutional layers and three layers of shortcut connections, and the three layers of shortcut connections are 1*1, 3*3 and 1*1 convolutions, respectively, wherein the 1*1 convolution layer is responsible for reducing and then restoring the size, making the 3*3 convolution layer a bottleneck with smaller input and output sizes,

[0014] Furthermore, the prediction module is composed of a state prediction subnet and a state update subnet, wherein the state prediction subnet estimates the state at the current time based on the posterior estimate at the previous time and obtains the prior estimate of the current time, and then uses the measurement value at the current time to correct the estimate of the prediction stage and obtain the posterior estimate of the current time.

[0015] Furthermore, the training process of the multi-scale detection network in step 2 is:

[0016] The size of the input image is set to 416*416*3, and it is input into the multi-scale detection network, and its features are extracted through the feature extraction network; the input image is divided into s*s grids, and each grid unit will predict three bounding boxes, and the bounding box with the largest intersection-over-union ratio with the labeled data will be used to predict the target.

[0017] Furthermore, the training process of the template matching network in step 2 is:

[0018] Select sample pairs from the ImageNet dataset, extract the target image and search image from two frames of the same video, where the size of the target image is 127*127*3 and the size of the search image is 255*255*3;

[0019] Then, the paired target image and search image are input into the subnet corresponding to the template matching network to obtain the target feature layer and the search feature layer, and the convolution results of the target feature layer and the search feature layer are matched with the true value of the label for similarity;

[0020] At the same time, the target loss function is used to calculate the error between the above matching result and the true value of the label, and the error is back-propagated to the input layer layer by layer;

[0021] Finally, the weights and biases in the network are adjusted based on the mini-batch stochastic gradient descent optimization algorithm to obtain the optimal error value, thus completing a network model training.

[0022] Repeat the above steps until the error value of the target loss function converges to the minimum value;

[0023] The formula for the bounding box coordinates is as follows:

[0024] b x =σ(t x )+c x (12)

[0025] b y =σ(t y )+c y (13)

[0026]

[0027]

[0028] Among them, (c x ,c y ) represents the coordinate of the upper left corner of the grid, (σ(t x ),σ(t y )) represents the offset, (p w , p h ) represents the width and height of the prior box, (b x ,b y ,b w ,b h ) represents the coordinates of the final predicted bounding box.

[0029] Furthermore, the specific process of step 3 is as follows:

[0030] (1) Inputting the original video sequence acquired by the camera into the detector of the trained target detection and tracking network model to obtain the detection result;

[0031] (2) Input the results of the above detectors into the intelligent selection strategy structure to obtain the target of interest;

[0032] (3) Inputting the object of interest obtained above into the target subnet of the template matching network, and inputting the next frame of the original video sequence into the search subnet of the template matching network;

[0033] (4) If the score of the above output result is less than the threshold, it is input into the prediction module, and its output result is input into the target subnet of the template matching network, and the next frame in the original video sequence is input into the search subnet of the template matching network;

[0034] (5) Repeat steps (3) to (4) to achieve intelligent selection and automatic real-time tracking of the target.

[0035] Furthermore, the process of the intelligent selection strategy is:

[0036] First, give the categories a weight. Different categories match different values. Assign higher values ​​to the categories you are interested in. Suppose we are interested in bottles, then the value assigned to bottles is the largest score among all categories. You can freely assign values ​​to each category according to different needs. Set the equation for the category score to

[0037]

[0038] where c(x i ) represents the value of the i-th target;

[0039] Sort the categories and assign scores to each category;

[0040] Among the multiple targets detected, since the area of ​​each target is different, the area of ​​each target is given a second weight. The product of the weight and the area of ​​the target area is the second factor in selecting the target. The size of the target area is set to

[0041] s i =ω i ·h i (3)

[0042] Among them, ω i Represents the width of the target bounding box, h i represents the height of the target bounding box,

[0043] The best target is selected, and the accuracy of the target is given a third weight. The product of the weight and the accuracy is the third factor of the target, indicating a higher accuracy for the target;

[0044] Through multi-scale detection of the first three frames, the speed of the target is given a fourth weight, indicating that the faster the target, the more interested it is, and the speed equation is set to

[0045]

[0046] Among them, d i is the distance from the first frame to the tenth frame, Δt represents the time from the first frame to the tenth frame, and is set to

[0047]

[0048] Δt=t2-t1 (6)

[0049] where x i The coordinates and time of the target in the first frame are (x1, y1) i and t1, the coordinates and time of the target in the tenth frame are (x2, y2) i and t2.

[0050] The present invention adopts the above technical solution, and the present invention has the following technical effects:

[0051] (1) The present invention proposes a novel target detection and tracking network model. Since the model uses a detector and tracker with good performance, the network model has the performance of real-time detection and automatic tracking;

[0052] (2) The present invention proposes an intelligent selection strategy, which can automatically acquire the target of interest and set corresponding parameters according to the needs, and has high flexibility and scalability;

[0053] (3) The present invention proposes a prediction module, which enables the present invention to have more accurate tracking results when the target is obscured. To this end, based on the above-mentioned innovations and advantages, the target detection and automatic tracking algorithm based on the intelligent selection strategy designed by the present invention has a high application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a network structure diagram of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clear, preferred embodiments are given to further describe the present invention in detail. However, it should be noted that many details listed in the specification are only for the purpose of enabling readers to have a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be implemented even without these specific details.

[0056] like Figure 1 As shown, the present invention provides a target detection and automatic tracking algorithm based on an intelligent selection strategy, and the tracking method comprises the following steps:

[0057] 1. Build target detection and tracking network model:

[0058] (1) The target detection and tracking network model constructed by the present invention is mainly composed of a multi-scale detection network, an intelligent selection strategy, a template matching network and a prediction module. Among them, the multi-scale detection network is mainly composed of a darknet-53 network and three prediction layers, which is used to extract features from the input video image and predict and locate the target.

[0059] (2) After the multi-scale detection network completes the prediction of the input video image, the present invention constructs an intelligent selection strategy to select the target. The principle of target selection of this strategy is composed of four factors, namely: target category, target area, target detection accuracy and target movement speed.

[0060] (3) After the target of interest is framed by the intelligent selection strategy, the present invention inputs the target of interest into the template matching network and tracks it in real time. The template matching network is composed of subnets, which share the same parameters, and the functions of the two subnets are to extract the features of the target image and the search image respectively.

[0061] (4) In a complex background, when the target is deformed, blocked, etc., the present invention uses a prediction module to predict the state and position of the target.

[0062] in:

[0063] The multi-scale detection network is mainly composed of a darknet-53 network and three prediction layers. The darknet-53 network has 53 convolutional layers, which are composed of 1*1 and 3*3 convolutional layers respectively. The performance of Darknet-53 is reflected in the difference in the basic units of the network. It uses ResNet as a skip layer connection method, and its performance is completely better than resnet-152 and resnet-101. Multi-scale detection effectively enhances the detection effect of objects of different sizes and blocked objects through the prediction method of three different scale prediction layers. At the same time, darknet-53 introduces a skip layer connection method to enhance the convergence effect, and adopts a random multi-scale training method to enhance robustness.

[0064] The intelligent selection strategy formula is shown in (1):

[0065] f(x)=max[ω1c i +ω2s i +ω3a i +ω4v i ] (1)

[0066] The intelligent selection strategy is mainly composed of four factors: target category, target area, target detection accuracy and target movement speed. Among them, the above four factors are respectively assigned four corresponding parameters, which are ω1, ω2, ω3 and ω4, and their respective parameter values ​​can be freely set according to specific requirements, with high flexibility and scalability. The intelligent selection strategy proposed by the present invention calculates the value of each target in the video image, and the target with the highest score is the target of interest selected by this strategy.

[0067] The principle of the intelligent selection strategy is as follows:

[0068] 1) First, give the category a weight. The present invention sets different categories to match different values, so we can assign higher values ​​according to the category of interest. For example, assuming we are interested in bottles, the value assigned to bottles is the largest score among all categories. Similarly, assuming the target we are interested in is cars, the value assigned to cars is the largest. Obviously, this method is highly flexible and can freely assign values ​​to each category according to different needs. The present invention sets the equation for the category score as

[0069]

[0070] where c(x i ) represents the value of the i-th target.

[0071] The present invention classifies the categories as follows:

[0072] The present invention sorts the categories and assigns scores to each category according to formula (2). Wherein, number represents the total number of categories. According to the COCO dataset, there are 80 categories in total, which are sorted starting from 0; category represents the category, and value represents the score of the category.

[0073] 2) Among the multiple targets detected by the multi-scale detection network, since the area of ​​each target is different, this paper assigns a second weight to the area of ​​each target. The product of the weight and the area of ​​the target area is the second factor in selecting the target, which means that we are more interested in large-area targets. The size of the target area is set to

[0074] s i =ω i ·h i (3)

[0075] Among them, ω i Represents the width of the target bounding box, h i Represents the height of the target bounding box.

[0076] 3) To further select the best target, we give a third weight to the accuracy of the target. The product of the weight and the accuracy is the third factor in our selection of the target to focus on, which indicates that we have a higher accuracy on the target.

[0077] 4) Through multi-scale detection of the first three frames, we can calculate the speed of each object, so we give the fourth weight to the speed of the object, which indicates that the faster the object is, the more interested we are in it.

[0078] The velocity equation is set as

[0079]

[0080] Among them, d i is the distance from the first frame to the tenth frame, Δt represents the time from the first frame to the tenth frame, and is set to

[0081]

[0082] Δt=t2-t1 (6)

[0083] where x i The coordinates and time of the target in the first frame are (x1, y1) i and t1, the coordinates and time of the target in the tenth frame are (x2, y2) i and t2.

[0084] The template matching network is composed of subnetworks, which are respectively called target subnetwork and search subnetwork, which extract features from target image and search image respectively, and share the same weights and biases. Among them, both the target subnetwork and the search subnetwork adopt the internal cropping (CIR) unit based on the "bottleneck" residual block. The CIR unit crops out the features affected by padding inside the block, thereby preventing the convolution filter from learning position deviation. Among them, the residual unit is the key module of the template matching network, which consists of 3 stacked convolutional layers and shortcut connections that bypass them. The three layers of shortcut connections are 1*1, 3*3 and 1*1 convolutions, respectively. The 1*1 convolution layer is responsible for reducing and then restoring the size, making the 3*3 convolution layer a bottleneck with smaller input and output sizes.

[0085] The prediction module is mainly composed of a state prediction subnet and a state update subnet. The state prediction subnet estimates the state at the current time based on the posterior estimate at the previous time and obtains the prior estimate at the current time. Then, the measurement value at the current time is used to correct the estimate in the prediction stage and obtain the posterior estimate at the current time.

[0086] The state prediction equation is set as

[0087]

[0088] in, represents the prior state estimate of the system last time, x t-1 represents the posterior state estimate of step t-1 when the observation value is known, representing the system's best prediction for the last time, λ t-1 represents the control quantity of the system, Q and P represent the system parameters.

[0089] The state update equation is set as

[0090]

[0091]

[0092]

[0093]

[0094] Among them, R t represents the gain, z t represents the measured value, H represents the matrix of the conversion state space, represents the covariance of the prior estimation error, and A t-1 represents the covariance of the post-estimation error at time t-1, B represents the covariance of the measurement noise, F represents the noise covariance, and I represents the unit matrix.

[0095] 2. Training target detection and tracking network model:

[0096] Since the multi-scale detection network and the template matching network are two modules with different functions and use different training sets, the present invention trains the multi-scale detection network and the template matching network separately.

[0097] 1. Training multi-scale detection network:

[0098] (1) First, the present invention sets the size of the input image to 416*416*3 and inputs it into the multi-scale detection network;

[0099] (2) Then, features are extracted through a feature extraction network;

[0100] (3) Then, the input image is divided into s*s grids, and each grid cell will predict three bounding boxes;

[0101] (4) Finally, the bounding box with the largest IOU with the ground truth is used to predict the target.

[0102] 2. Training template matching network:

[0103] (1) First, the present invention selects sample pairs from the ImageNet dataset and extracts the target image and the search image from two frames of the same video. The size of the target image is 127*127*3, and the size of the search image is 255*255*3;

[0104] (2) Then, the paired target image and search image are input into the corresponding subnet of the template matching network to obtain the target feature layer and the search feature layer. The convolution results of the target feature layer and the search feature layer are matched with the true value of the label for similarity;

[0105] (3) At the same time, the target loss function is used to calculate the error between the matching result and the true value of the label, and the error is back-propagated to the input layer layer by layer;

[0106] (4) Finally, the weights and biases in the network are adjusted based on the mini-batch stochastic gradient descent optimization algorithm to obtain the optimal error value, thereby completing a network model training;

[0107] (5) Repeat the above steps until the error value of the objective loss function converges to the minimum value.

[0108] in:

[0109] The formula for the bounding box coordinates is as follows:

[0110] b x =σ(t x )+c x (12)

[0111] b y =σ(t y )+c y (13)

[0112]

[0113]

[0114] Among them, (c x ,c y ) represents the coordinate of the upper left corner of the grid, (σ(t x ),σ(t y )) represents the offset, (p w , p h ) represents the width and height of the prior box, (b x ,b y ,b w ,b h ) represents the coordinates of the final predicted bounding box.

[0115] 3. Test target detection and tracking network model:

[0116] Use the above trained target detection and tracking network model to perform target tracking test. The specific steps are as follows:

[0117] (1) inputting the original video sequence acquired by the camera into the detector of the target detection and tracking network model that has been trained by the present invention to obtain the detection result;

[0118] (2) Input the results of the above detectors into the intelligent selection strategy structure to obtain the target of interest;

[0119] (3) Inputting the object of interest obtained above into the target subnet of the template matching network, and inputting the next frame of the original video sequence into the search subnet of the template matching network;

[0120] (4) If the score of the above output result is less than the threshold, it is input into the prediction module, and its output result is input into the target subnet of the template matching network, and the next frame in the original video sequence is input into the search subnet of the template matching network;

[0121] (5) Repeat steps (3) to (4) above to achieve intelligent selection and automatic real-time tracking of the target.

[0122] The present invention proposes a target detection and automatic tracking algorithm based on an intelligent selection strategy. The method consists of a multi-scale detection network, an intelligent selection strategy, a template matching network and a prediction module. In this paper, a multi-scale detection network is first used to detect the target in real time, and then the detection candidate frame is selected according to the intelligent selection strategy. Different influencing factors can be set according to different needs, and it has high flexibility and scalability. Finally, the target of interest is tracked in real time through the template matching network. At the same time, in order to solve the problem of the target being blocked or temporarily disappearing, the present invention adds a prediction module. When the target is blocked in uniform motion, the algorithm can predict the state and position of the target at the next few moments, effectively solving the problem of the target being blocked and temporarily disappearing. Quantitative evaluation tests show that under complex background environments, the algorithm proposed in the present invention realizes the functions of real-time detection, intelligent selection and automatic tracking of targets, and has a good promoting effect and reference value for the development of the field of computer vision.

[0123] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A target detection and automatic tracking method based on intelligent selection strategy, characterized in that: The method comprises the following steps: Step 1: Build a target detection and tracking network model, which consists of a multi-scale detection network, an intelligent selection strategy, a template matching network, and a prediction module; Step 2: Train the target detection and tracking network model. Train the multi-scale detection network and template matching network on the COCO dataset and ImageNet dataset respectively. Step 3: Test the target detection and tracking network model. Use the trained target detection and tracking network model to perform target detection and tracking tests to achieve intelligent selection and automatic real-time tracking of targets. In the step 1, the multi-scale detection network is composed of a darknet-53 network and three prediction layers, which are used to extract features of the input video image and predict and locate the target. After the multi-scale detection network completes the prediction of the input video image, the intelligent selection strategy selects the target. The principle of selecting the target by the intelligent selection strategy is composed of four factors, which are: target category, target area, target detection accuracy and target movement speed. After the intelligent selection strategy frames the target of interest, the target of interest is input to the template matching network and tracked in real time. The template matching network is composed of two subnets, which share the same parameters, and the functions of the two subnets are to extract features of the target image and the search image respectively. In a complex background, when the target is deformed or blocked, the prediction module predicts the state and position of the target; The specific process of step 3 is as follows: (1) Inputting the original video sequence acquired by the camera into the detector of the trained target detection and tracking network model to obtain the detection result; (2) Input the results of the above detectors into the intelligent selection strategy structure to obtain the target of interest; (3) Inputting the object of interest obtained above into the target subnet of the template matching network, and inputting the next frame of the original video sequence into the search subnet of the template matching network; (4) If the score of the above output result is less than the threshold, it is input into the prediction module, and its output result is input into the target subnet of the template matching network, and the next frame in the original video sequence is input into the search subnet of the template matching network; (5) Repeat steps (3) to (4) to achieve intelligent selection and automatic real-time tracking of the target.

2. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1, characterized in that: The darknet-53 network has a total of 53 convolutional layers, which are composed of 1*1 and 3*3 convolutional layers respectively. The performance of Darknet-53 is reflected in the difference of the basic units of the network. It adopts ResNet as the skip layer connection method. Multi-scale detection enhances the detection of objects of different sizes and blocked objects through the prediction method of three different scale prediction layers. At the same time, darknet-53 introduces the skip layer connection method to enhance convergence, and adopts the random multi-scale training method to enhance robustness.

3. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1, characterized in that: The selection principle of the intelligent selection strategy is as follows: f(x)=max[ω1c i +ω2s i +ω3a i +ω4v i ] (1) Among them, c i is the target category, s i is the target area, a i is the target detection accuracy, v i The above four factors are assigned four corresponding parameters, namely ω1, ω2, ω3 and ω4, which can freely set their own parameter values ​​according to specific requirements. It has high flexibility and scalability. The intelligent selection strategy calculates the value of each target in the video image, and the target with the highest score is the target of interest selected by this strategy.

4. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1, characterized in that: The template matching network is composed of subnets, which include a target subnet and a search subnet. The target image and the search image are respectively extracted with features, and the same weights and biases are shared. The target subnet and the search subnet both use internal cropping units based on bottleneck residual blocks. The internal cropping units crop features affected by zero padding inside the block to prevent the convolution filter from learning position deviations. The residual unit is a key module of the template matching network, which is composed of three stacked convolutional layers and three layers of shortcut connections. The three layers of shortcut connections are 1*1, 3*3 and 1*1 convolutions, respectively. The 1*1 convolution layer is responsible for reducing and then restoring the size, making the 3*3 convolution layer a bottleneck with smaller input and output sizes.

5. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1, characterized in that: The prediction module consists of a state prediction subnet and a state update subnet, wherein the state prediction subnet estimates the state at the current time based on the posterior estimate at the previous time and obtains the prior estimate at the current time, and then uses the measured value at the current time to correct the estimate of the prediction stage and obtain the posterior estimate at the current time.

6. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1, characterized in that: The training process of the multi-scale detection network in step 2 is: The size of the input image is set to 416*416*3, and it is input into the multi-scale detection network, and features are extracted through the feature extraction network; The input image is divided into s*s grids, and each grid unit will predict three bounding boxes, and the bounding box with the largest intersection-over-union ratio with the labeled data will be used to predict the target.

7. The target detection and automatic tracking method based on intelligent selection strategy according to claim 1 is characterized in that: The training process of the template matching network in step 2 is: Select sample pairs from the ImageNet dataset, extract the target image and search image from two frames of the same video, where the size of the target image is 127*127*3 and the size of the search image is 255*255*3; Then, the paired target image and search image are input into the subnet corresponding to the template matching network to obtain the target feature layer and the search feature layer, and the convolution results of the target feature layer and the search feature layer are matched with the true value of the label for similarity; At the same time, the target loss function is used to calculate the error between the above matching result and the true value of the label, and the error is back-propagated to the input layer layer by layer; Finally, the weights and biases in the network are adjusted based on the mini-batch stochastic gradient descent optimization algorithm to obtain the optimal error value, thus completing a network model training. Repeat the above steps until the error value of the target loss function converges to the minimum value; The formula for the bounding box coordinates is as follows: b x =σ(t x )+c x (12) b y =σ(t y )+c y (13) Among them, (c x ,c y ) represents the coordinate of the upper left corner of the grid, (σ(t x ),σ(t y )) represents the offset, (p w , p h ) represents the width and height of the prior box, (b x ,b y ,b w ,b h ) represents the coordinates of the final predicted bounding box.

8. The target detection and automatic tracking method based on intelligent selection strategy according to claim 3 is characterized in that: The process of the intelligent selection strategy is: First, give the categories a weight. Different categories match different values. Assign higher values ​​to the categories you are interested in. Suppose we are interested in bottles, then the value assigned to bottles is the largest score among all categories. You can freely assign values ​​to each category according to different needs. Set the equation for the category score to where c(x i ) represents the value of the i-th target; Sort the categories and assign scores to each category; Among the multiple targets detected, since the area of ​​each target is different, the area of ​​each target is given a second weight. The product of the weight and the area of ​​the target area is the second factor in selecting the target. The size of the target area is set to s i =ω i ·h i (3) Among them, ω i Represents the width of the target bounding box, h i represents the height of the target bounding box, The best target is selected, and the accuracy of the target is given a third weight. The product of the weight and the accuracy is the third factor of the target, indicating a higher accuracy for the target; Through multi-scale detection of the first three frames, the speed of the target is given a fourth weight, indicating that the faster the target, the more interested it is, and the speed equation is set to Among them, d i is the distance from the first frame to the tenth frame, Δt represents the time from the first frame to the tenth frame, and is set to Δt=t2-t1 (6) where x i The coordinates and time of the target in the first frame are (x1, y1)i and t1, and the coordinates and time of the target in the tenth frame are (x2, y2)i and t2.

Images