An efficient pedestrian following method that can resist long-term occlusion

Abstract

The present invention discloses an efficient pedestrian following method resistant to long-term occlusion, comprising two parts: constructing a behavior discriminator based on pedestrian re-identification for target tracking data and a novel pedestrian feature dynamic update mechanism. The main steps of the method are: obtaining target detection information of selected image frames to initialize a pedestrian tracker, constructing a behavior discriminator to supervise the changes in motion states between pedestrians, predicting target information under continuous observation states by associating inter-frame trajectories of the tracker, and restoring target identity representation under discrete observation point states based on a pedestrian re-identification network; dynamically adjusting feature update weights of historical frames and current frames according to the degree of target occlusion, weakening noise interference introduced by occlusion, and optimizing pedestrian following effects; the present invention is applicable to a variety of tracking algorithms to improve the performance against long-term occlusion, and the tracking stability and computational cost are superior to the current method of tracking and then re-identifying pedestrians frame by frame, and has been successfully applied in robot following tasks.

Description

Technical Field

[0001] The present invention belongs to the field of robot vision in pedestrian motion scenes, and in particular relates to a pedestrian following method that is efficient and resistant to long-term occlusion. Background Art

[0002] With the development of technologies such as unmanned vehicles / drones and humanoid robots, how to achieve reliable long-term positioning and following of targets is an important part of the application of robot vision. Especially in pedestrian motion scenes, the robot's ability to follow the target pedestrian is the key to achieving service and collaborative robot tasks. However, due to challenges such as dynamic occlusion, changes in pedestrian appearance, and multi-target interference in complex scenes, the existing pedestrian following methods still do not perform well in long-term tracking tasks.

[0003] The difficulty of pedestrian following tasks lies in how to effectively distinguish target pedestrians in a multi-target environment while maintaining stable tracking of the target under dynamic occlusion. Traditional methods mainly rely on motion information and the initial appearance features of the target, and use Kalman filtering or Hungarian algorithm to associate trajectories, but are prone to losing the target when faced with long-term occlusion or drastic changes in appearance. After the introduction of deep learning methods, the ability to identify the identity of the target pedestrian is enhanced through pedestrian re-identification technology, thereby improving the robustness of tracking. However, existing methods still face the following problems: (1) Existing pedestrian following methods often use frame-by-frame pedestrian re-identification after tracking, and add a pedestrian posture detection network during the tracking process. It may be difficult to achieve real-time effects when fully deployed on some edge devices with limited computing resources; (2) When occlusion occurs, the appearance information of other targets may be mixed into the pedestrian frame, resulting in contamination of the target feature update, which is prone to mis-following or loss in complex scenes, thus affecting the following effect.

[0004] Therefore, an efficient pedestrian following method that is resistant to long-term occlusion is needed, which can effectively distinguish target pedestrians in complex dynamic environments, solve the feature pollution problem caused by occlusion, maintain stable tracking of target pedestrians, and improve the model's anti-interference ability in multi-target scenarios. Summary of the invention

[0005] The purpose of the present invention is to provide a pedestrian following method that is efficient and resistant to long-term occlusion.

[0006] The technical solution to achieve the purpose of the present invention is: an efficient pedestrian following method that resists long-term occlusion, comprising the following steps:

[0007] Obtain RGB image data from the camera sensor, input it into the target detection network, use non-maximum suppression to generate target detection information, and obtain pedestrian detection frames and pedestrian confidence through category screening;

[0008] Initialize the pedestrian tracker based on the target detection information of the selected image frame, and build a behavior discriminator for pedestrian re-identification to supervise the motion state changes between pedestrians;

[0009] Based on the constructed behavior discriminator for pedestrian re-identification, it is determined that the tracker is scheduled under continuous observation to predict and associate the pedestrian's motion trajectory, and the target pedestrian identity representation is obtained after filtering and updating through the predicted trajectory and observation information. And the target pedestrian tracking box ; Determine the scheduling of the pedestrian re-identification network to restore the target pedestrian identity representation under discrete observation node states And the target pedestrian tracking frame ;

[0010] According to the occlusion degree of the moving target, weights are dynamically assigned to the target pedestrian features in the historical observation process and the current target pedestrian features, and an anti-occlusion feature update function module is established.

[0011] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.

[0012] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0013] A computer program product comprises a computer program, which implements the steps of the above method when executed by a processor.

[0014] Compared with the prior art, the present invention has the following significant advantages: (1) The present invention proposes a time-sharing pedestrian re-identification strategy, which embeds the pedestrian re-identification mechanism into the discrete nodes of the tracking method process. Compared with the traditional pure tracking method after detection, the introduction of the pedestrian re-identification mechanism can greatly improve the stability and accuracy of pedestrian following. Compared with the existing method of using pedestrian re-identification frame by frame after the two-stage detection-tracking, the time-sharing pedestrian re-identification strategy of the present invention can reduce the consumption of computing resources, and is more efficient on the basis of ensuring accuracy and stability, which is conducive to the deployment of edge devices; (2) The present invention proposes an anti-occlusion dynamic pedestrian feature update module, so that the feature update weight can be adaptively adjusted according to the degree of target occlusion. In the pedestrian motion scene, it can reduce the accumulation of feature pollution caused by the target pedestrian being occluded by other pedestrians, thereby improving the accuracy of pedestrian re-identification; (3) The behavior discriminator module proposed in the present invention realizes an end-to-end process, which can be applied to a variety of multi-target tracking algorithms to improve the target following effect. By selecting different target category detection methods, it can also be applied to other target category data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the pedestrian following method based on pedestrian re-identification in a moving scene that is efficient and resistant to long-term occlusion.

[0016] Figure 2 It is a schematic diagram of the determination processing process of the behavior discriminator based on pedestrian re-identification of the present invention.

[0017] Figure 3 Schematic diagram of the weight allocation strategy process of the anti-occlusion dynamic pedestrian feature update module of the present invention.

[0018] Figure 4 Schematic diagram of the comparison between the present invention and the traditional method in terms of pedestrian following evaluation indicators. DETAILED DESCRIPTION

[0019] like Figure 1 As shown, the pedestrian following method based on pedestrian re-identification in a moving scene with high efficiency and resistance to long-term occlusion of the present invention comprises the following steps:

[0020] (1) Construct a behavior discriminator based on pedestrian re-identification for target tracking data and supervise the motion state changes between pedestrians to achieve identity consistency representation and tracking trajectory maintenance of target pedestrians under long-term occlusion:

[0021] Step 1: Get RGB image data from the camera sensor After being input into the target detection network, non-maximum suppression is used to generate target detection information, and pedestrian detection boxes are obtained through category screening. and pedestrian confidence .

[0022] Step 2: Initialize the pedestrian tracker based on the target detection information of the selected image frame , and construct a behavior discriminator for pedestrian re-identification to supervise the changes in motion states between pedestrians.

[0023] Step 3: Based on the constructed behavior discriminator for pedestrian re-identification, determine the scheduling tracker under continuous observation to predict and associate pedestrian motion trajectories, and obtain the filtered and updated target pedestrian identity representation through the predicted trajectory and observation information. And the target pedestrian tracking box ; Determine the scheduling of the pedestrian re-identification network to restore the target pedestrian identity representation under discrete observation node states And the target pedestrian tracking frame .

[0024] (2) Construct a new pedestrian feature dynamic update mechanism:

[0025] Step 4: According to the degree of occlusion of the moving target, dynamically assign weights to the target pedestrian features in the historical observation process and the current target pedestrian features, and establish an anti-occlusion feature update function module.

[0026] Furthermore, first, the target detection information of the selected image frame is obtained to initialize the pedestrian tracker. Then, a behavior discriminator is constructed to supervise the changes in the motion state between pedestrians. Then, the target motion state is supervised by the behavior discriminator to determine whether to use the tracker to associate the inter-frame trajectory to predict the target information under the continuous observation state or to restore the target identity under the discrete observation point state based on the pedestrian re-identification network. Among them, an anti-occlusion dynamic pedestrian feature update function module is introduced under the continuous observable state, and the feature update weights of the historical frame and the current frame are dynamically adjusted according to the target occlusion degree, so as to weaken the noise interference introduced by the occlusion, and obtain the following information of the target pedestrian in the final pedestrian motion scene.

[0027] According to the target-oriented tracking data, a behavior discriminator based on pedestrian re-identification is constructed and the motion state changes between pedestrians are supervised, so as to achieve the identity consistency representation and tracking trajectory maintenance of the target pedestrian under long-term occlusion.

[0028] Step 1: Get RGB image data from the camera sensor After being input into the target detection network, non-maximum suppression is used to generate target detection information, and pedestrian detection boxes are obtained through category screening. and pedestrian confidence .

[0029] Pedestrian following method based on target detection is an important task in the field of robot vision. In the scene of pedestrian movement, unmanned vehicles / machines and robots need RGB image data obtained by camera sensors. As input. First, the initial detection information is obtained by using the lightweight network of yolo target detection, and then the filtered target detection information is obtained by using non-maximum suppression , get the pedestrian detection frame by category screening and pedestrian confidence ,in , , , Indicates the number of target detection boxes, Indicates the number of pedestrian detection boxes, Indicates target detection information , Indicates pedestrian detection box information , Indicates the coordinates of the upper left corner of the detection box. Indicates the width and height of the detection frame. Represents the confidence of the detection box, Indicates the target category.

[0030] Step 2: Initialize the pedestrian tracker based on the target detection information of the selected image frame , and construct a behavior discriminator for pedestrian re-identification to supervise the changes in motion states between pedestrians.

[0031] (1) The pedestrian tracker is constructed and associated with the pedestrian trajectory matching based on the appearance branch and the motion branch. The appearance branch uses a lightweight appearance feature extractor, and the formula is as follows:

[0032]

[0033]

[0034] in For the Frame image Pedestrian detection box information, is the image capture function, is the appearance feature extraction network, For the Frame image Trajectory pedestrian features; To So far track The characteristics of pedestrians, is the weight coefficient;

[0035] The motion branch uses a nonlinear Kalman filter to predict and update the trajectory of pedestrian motion. The core of the Kalman gain is to achieve nonlinear filter adaptation and dynamically balance the prediction and observation information. , the formula is as follows:

[0036]

[0037]

[0038] in, is the time step The forecast covariance under , which represents the uncertainty of the current state, is the observation model, which is used to map the state from the estimation space to the observation space. is the preset constant measurement noise covariance, is the time step The confidence of each detection box is is the noise covariance that is dynamically adjusted based on the confidence level. Replace with variable , nonlinear filtering effect can be achieved;

[0039] (2) Constructing a behavior discriminator for person re-identification , based on the changes in pedestrian motion scenes, it can be divided into three observation states, namely, the target continuous observable state , the target is continuously unobservable , observation state at discrete nodes , the decision logic is expressed as follows:

[0040]

[0041] in, Indicates whether the input pedestrian identity is in an observable state. Indicates to The identity set of target pedestrians that appeared before the frame, Indicates that the tracker updates the output The set of pedestrian identities in the frame, Indicates to The set of pedestrian identities output by the frame-ahead tracker update, To take the large function.

[0042] Furthermore, the behavior discriminator based on the constructed pedestrian re-identification needs to determine the motion state of the current target pedestrian based on the historical tracking update output information and the current frame tracking update output information. Under continuous observation, if the target is in a continuously observable state, the behavior discriminator will give the tracker of the current target pedestrian a larger expectation value without enabling the pedestrian re-identification mechanism. The identity representation and tracking frame of the target pedestrian are updated and output by the tracker; if the target is in a continuously unobservable state, it is determined that the target is blocked for a long time or has left the current camera field of view for a long time. The behavior discriminator will make full use of the tracker to supervise the motion state of other pedestrians. At the same time, the identity of the target pedestrian is set to -1 by default, and the tracking box of the target pedestrian is set to empty. At this time, there is no need to start pedestrian re-identification to occupy additional computing resources; if the target is in the observation state at a discrete node, it means that the target pedestrian has changed from an unobservable state to an observable state. This short process can be considered as a discrete node distributed on the time axis of the entire pedestrian motion scene. The behavior determiner will enable the pedestrian re-identification mechanism, re-identify the target pedestrian and re-associate its trajectory through the similarity measurement of the cosine distance, and update the identity representation of the target pedestrian to achieve a long-term and stable pedestrian following effect that is resistant to long-term occlusion.

[0043] Step 3: Based on the constructed behavior discriminator for pedestrian re-identification, determine the scheduling tracker under continuous observation to predict and associate pedestrian motion trajectories, and obtain the filtered and updated target pedestrian identity representation through the predicted trajectory and observation information. And the target pedestrian tracking box ; Determine the scheduling of the pedestrian re-identification network to restore the target pedestrian identity representation under discrete observation node states And the target pedestrian tracking frame .

[0044] (1) Update the target pedestrian information in a continuously observable state, including the motion scene where the selected target pedestrian appears within the camera's current observable field of view, and the target detection network anchors the detection frame of the selected target pedestrian, as follows:

[0045]

[0046]

[0047] in, Represents the tracker update function, receiving and processing the current Frame detection information, Indicates Pedestrian detection frame information of the frame, Indicates The corresponding pedestrian detection box confidence of the frame, when the observation state When the tracker update output contains the target pedestrian information, the behavior determiner Determine the observed state output, represents the target pedestrian information extraction function, which is used to output the target pedestrian identity representation and the target pedestrian tracking box;

[0048]

[0049] in, It is a function module for updating the anti-occlusion dynamic pedestrian features in the pedestrian following task. Indicates to follow The smooth features of the target pedestrian during the period, Indicates the current Frame target pedestrian features.

[0050] (2) Retain the target pedestrian information in the continuous unobservable state, including the motion scenes where the selected target pedestrian is blocked by other pedestrians in the field of view for a long time, resulting in the failure of tracking trajectory; the selected target pedestrian is blocked by other objects in the field of view for a long time, resulting in the failure of tracking trajectory; the selected target pedestrian leaves the current observation field of view for a long time, resulting in the failure of tracking trajectory, as follows:

[0051]

[0052]

[0053]

[0054] Among them, when the observation state When the tracker update output does not contain the target pedestrian information, the target pedestrian identity is represented by ;

[0055] (3) Update the target pedestrian information in the discrete observation node state, including the motion scene After observing the scene, the selected target pedestrian reappears within the camera's current observable field of view, and the target detection network anchors the detection frame for the selected target pedestrian, as follows:

[0056]

[0057]

[0058]

[0059] in, represents the pedestrian re-identification network. When the behavior determiner determines that the observation state is When the target pedestrian identity is restored based on the pedestrian re-identification network And the target pedestrian tracking frame , activates new tracks of the target pedestrian in the camera’s field of view.

[0060] The entire judgment process is shown in the attached Figure 2 As shown. First, the tracker needs to be initialized based on the initial selected target detection information , using target tracking data to build a behavior discriminator based on pedestrian re-identification Then use the behavior discriminator to , the tracking result of the detection information of the frame is used to determine the state, and the continuous unobservable state can be distinguished according to whether a pedestrian appears , distinguishing continuous observable states according to whether the pedestrian re-identification mechanism is used and discrete node observation states Then, for different motion scenes, the behavior determiner activates different mechanisms to determine the identity information and tracking frame information of the target pedestrian. Finally, the following information of the target pedestrian is obtained to achieve an efficient pedestrian following effect that is resistant to long-term occlusion.

[0061] Furthermore, since the target pedestrian is occluded to varying degrees by other pedestrians during the movement process, the intercepted target pedestrian tracking box image features will be mixed with noise feature pollution. This noise pollution will gradually accumulate in the traditional feature update mechanism as the number of occlusions increases, and will have a certain impact on the stage of enabling the pedestrian re-identification mechanism. Therefore, a more robust target pedestrian feature update mechanism is needed to achieve anti-occlusion effect.

[0062] Step 4: According to the degree of occlusion of the moving target, dynamically assign weights to the target pedestrian features in the historical observation process and the current target pedestrian features, and establish an anti-occlusion feature update function module.

[0063] The process of pedestrian feature weight allocation strategy based on different occlusion levels is shown in the attached figure. Figure 3 As shown. The historical target pedestrian features and the current frame target pedestrian features are balanced by updating the weights according to the dynamic features. The value of the feature update weight is affected by two factors, namely the number of other pedestrian frames that overlap with the target pedestrian frame and the sum of the areas of the overlapping parts with the target frame. The update process is enabled under the determined continuous observable state. The specific formula is as follows:

[0064]

[0065]

[0066] ,

[0067] in, is the overlapping area between other pedestrian frames and the target pedestrian frame, is the number of other pedestrian frames that overlap with the target pedestrian frame, is the sum of the areas of the parts overlapping with the target pedestrian frame, is the area of ​​the target pedestrian box, is the adaptive weight coefficient of the occlusion degree, is the adjustment factor to control the occlusion The impact of ,default , Is the basic weight coefficient, default , To take the small function.

[0068] The application of the present invention requires an RGB video sequence data set or a real-time camera data stream in a pedestrian motion scene to test the effect of the above-mentioned anti-long-term occlusion pedestrian following method, such as Figure 4 As shown, the public pedestrian following dataset used in the present invention is proposed in ICVS, where ICVS stands for the International Conference on Computer Vision Systems. The dataset contains 11 video sequences, a total of more than 56,000 RGB images, and has a variety of challenging scenes, including posture changes, strong lighting changes, appearance changes, partial and complete occlusions, etc.; in addition, in order to further test the anti-long-term occlusion performance, the self-annotated dataset used in the present invention contains more complex long-term occlusion scenes.

[0069] The present invention can be applied to a variety of multi-target tracking algorithms. The single-target tracking types include SiamRPN++ and STARK, where SiamRPN++ represents an advanced twin network region proposal network and STARK represents a space-time attention network tracker; the multi-target tracking types include SORT, OC-SORT, ByteTrack and Deepsort, where SORT represents simple online real-time tracking, OC-SORT represents high-order center point online real-time tracking, ByteTrack represents low-confidence target tracking based on detection results, and Deepsort represents simple online real-time tracking with appearance features; the robot pedestrian following types include SORT w / OCL_REID, OC-SORT w / OCL_REID, ByteTrack w / OCL_REID, Deepsort w / osnet, Strongsort w / osnet, SORT_ours, OC-SORT_ours, ByteTrack_ours, Deepsort_ours and Strongsort_ours, where SORT w / OCL_REID represents simple online real-time tracking introducing continuous online learning of pedestrian re-identification, OC-SORT w / OCL_REID represents the introduction of high-order center point online real-time tracking of continuous online learning for pedestrian re-identification, ByteTrack w / OCL_REID represents the introduction of low-confidence target tracking based on detection results for continuous online learning for pedestrian re-identification, Deepsort w / osnet represents the introduction of simple online real-time tracking with appearance features of a cross-scale feature fusion network, Strongsort w / osnet represents the introduction of online real-time tracking of an enhanced appearance model of a cross-scale feature fusion network, SORT_ours represents the introduction of the simple online real-time tracking of the present invention, OC-SORT_ours represents the introduction of high-order center point online real-time tracking of the present invention, ByteTrack_ours represents the introduction of low-confidence target tracking based on detection results of the present invention, Deepsort_ours represents the introduction of simple online real-time tracking with appearance features of the present invention, and Strongsort_ours represents the introduction of online real-time tracking of the enhanced appearance model of the present invention.

[0070] Tests on the public pedestrian following dataset icvs and a personal annotated dataset containing more complex scenes such as long-term occlusions show that this efficient pedestrian following method that is resistant to long-term occlusions is superior to current pedestrian following methods in terms of experimental evaluation indicators and actual visualization effects.

Claims

1. An efficient pedestrian following method that can resist long-term occlusion, characterized in that: The following steps are involved: Obtain RGB image data from the camera sensor, input it into the target detection network, use non-maximum suppression to generate target detection information, and obtain pedestrian detection frames and pedestrian confidence through category screening; Initialize pedestrian tracker based on target detection information in selected image frames , and construct a behavior discriminator for pedestrian re-identification to supervise the changes in the motion state between pedestrians, as follows: (1) The pedestrian tracker is constructed and associated with the pedestrian trajectory matching based on the appearance branch and the motion branch. The appearance branch uses a lightweight appearance feature extractor, and the formula is as follows: ； ； in For the Frame image Pedestrian detection box information, is the image capture function, is the appearance feature extraction network, For the Frame image Trajectory pedestrian features; To So far track The characteristics of pedestrians, is the weight coefficient; The motion branch uses a nonlinear Kalman filter to predict and update the trajectory of pedestrian motion. The core of the Kalman gain is to achieve nonlinear filter adaptation and dynamically balance the prediction and observation information. , the formula is as follows: ； ； in, is the time step The forecast covariance under , which represents the uncertainty of the current state, is the observation model, which is used to map the state from the estimation space to the observation space. is the preset constant measurement noise covariance, is the time step The confidence of each detection box is is the noise covariance that is dynamically adjusted based on the confidence level. Replace with variable , to achieve nonlinear filtering effect; (2) Constructing a behavior discriminator based on pedestrian re-identification , based on the changes in pedestrian motion scenes, it can be divided into three observation states, namely, the target continuous observable state , the target is continuously unobservable , observation state at discrete nodes , the decision logic is expressed as follows: ； in, Indicates whether the input pedestrian identity is in an observable state. Indicates to The identity set of target pedestrians that appeared before the frame, Indicates that the tracker updates the output The set of pedestrian identities in the frame, Indicates to The set of pedestrian identities output by the frame-ahead tracker update, To take the largest function; Based on the constructed behavior discriminator for pedestrian re-identification, it is determined that the tracker is scheduled under continuous observation to predict and associate the pedestrian's motion trajectory, and the target pedestrian identity representation is obtained after filtering and updating through the predicted trajectory and observation information. and target pedestrian tracking box ; Determine the scheduling of the pedestrian re-identification network to restore the target pedestrian identity representation under discrete observation node states And the target pedestrian tracking frame , the process includes the following three parts: (1) Update the target pedestrian information in a continuously observable state, including the motion scene where the selected target pedestrian appears within the camera's current observable field of view, and the target detection network anchors the detection frame of the selected target pedestrian, as follows: ； ； in, Represents the tracker update function, receiving and processing the current Frame detection information, Indicates Pedestrian detection frame information of the frame, Indicates The corresponding pedestrian detection box confidence of the frame, when the observation state When the tracker update output contains the target pedestrian information, the behavior determiner Determine the observed state output, represents the target pedestrian information extraction function, which is used to output the target pedestrian identity representation and the target pedestrian tracking box; ； in, It is a function module for updating the anti-occlusion dynamic pedestrian features in the pedestrian following task. Indicates to follow The smooth features of the target pedestrian during the period, Indicates the current Frame target pedestrian features; (2) Retain the target pedestrian information in the continuous unobservable state, including the motion scenes where the selected target pedestrian is blocked by other pedestrians in the field of view for a long time, resulting in the failure of tracking trajectory; the selected target pedestrian is blocked by other objects in the field of view for a long time, resulting in the failure of tracking trajectory; the selected target pedestrian leaves the current observation field of view for a long time, resulting in the failure of tracking trajectory, as follows: ； ； ； Among them, when the observation state When the tracker update output does not contain the target pedestrian information, the target pedestrian identity is represented by ; (3) Update the target pedestrian information in the discrete observation node state, including the motion scene After observing the scene, the selected target pedestrian reappears within the camera's current observable field of view, and the target detection network anchors the detection frame for the selected target pedestrian, as follows: ； ； ； in, represents the pedestrian re-identification network. When the behavior determiner determines that the observation state is When the target pedestrian identity is restored based on the pedestrian re-identification network And the target pedestrian tracking frame , activate the new trajectory of the target pedestrian in the camera field of view; According to the occlusion degree of the moving target, weights are dynamically assigned to the target pedestrian features in the historical observation process and the current target pedestrian features, and an anti-occlusion feature update function module is established.

2. The efficient pedestrian following method against long-term occlusion according to claim 1 is characterized in that: Get RGB image data from the camera sensor , after being input into the target detection network, non-maximum suppression is used to generate target detection information , get the pedestrian detection frame by category screening and pedestrian confidence , as follows: First, we need the RGB image data obtained by the camera sensor As input, the initial detection information is obtained by using the lightweight network of yolo target detection, and then the filtered target detection information is obtained using non-maximum suppression , get the pedestrian detection frame by category screening and pedestrian confidence ,in , , , Indicates the number of target detection boxes, Indicates the number of pedestrian detection boxes, Indicates target detection information , Indicates pedestrian detection box information , Indicates the coordinates of the upper left corner of the detection box. Indicates the width and height of the detection frame. Represents the confidence of the detection box, Indicates the target category.

3. The efficient pedestrian following method against long-term occlusion according to claim 1 is characterized in that: According to the degree of occlusion of the moving target, weights are dynamically assigned to the target pedestrian features in the historical observation process and the current target pedestrian features, and an anti-occlusion feature update function is established, as follows: ； ； ； in, is the overlapping area between other pedestrian frames and the target pedestrian frame, is the number of other pedestrian frames that overlap with the target pedestrian frame, is the sum of the areas of the parts overlapping with the target pedestrian frame, is the area of ​​the target pedestrian box, is the adaptive weight coefficient of the occlusion degree, is the adjustment factor to control the occlusion The magnitude of the impact, , is the basic weight coefficient, To take the small function.

4. The efficient pedestrian following method against long-term occlusion according to claim 3 is characterized in that: ， 。 5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images