Pedestrian trajectory prediction method of lightweight interactive perception GAN
By using a lightweight interactive perception GAN model, the problems of model complexity, deployment difficulties, and high trajectory repetition rate in pedestrian trajectory prediction in closed scenarios are solved. This enables efficient and accurate pedestrian trajectory prediction on edge devices, meeting the operational management and security control needs in closed scenarios.
Patent Information
- Application Number
- CN202511514498.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2026-02-17
AI Technical Summary
Existing technologies for predicting pedestrian trajectories in closed environments suffer from problems such as complex models, difficulty in deploying on edge devices, high trajectory repetition rates, long training cycles, and failure to fully consider the interaction between pedestrians and surrounding objects, resulting in large prediction errors and failing to meet the needs of accurate passenger flow guidance and safety control.
A lightweight interactive perception GAN model is adopted. By combining a GRU temporal encoder and an interaction feature fusion layer with a dual-constraint discriminator, a lightweight GAN model is constructed. It uses edge devices to predict pedestrian trajectories, combines distance weights and interaction object features to realize the quantitative expression of interaction relationships, and generates efficient and accurate trajectory predictions through training with a cross-entropy loss function.
It enables efficient deployment of pedestrian trajectory prediction on edge devices, reduces model complexity and trajectory repetition rate, improves prediction accuracy and adaptability, and meets the practical needs of closed scenarios.
Smart Images
Figure CN121545112A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer vision and deep learning, in particular to a lightweight interactive perception GAN (Generative Adversarial Network) pedestrian trajectory prediction method suitable for closed scenes, which can be widely applied to pedestrian behavior analysis scenes in closed scenes such as shopping mall passenger flow guidance, community security monitoring, supermarket flow planning, etc., providing data support for scene operation management and safety prevention and control. BACKGROUND
[0002] In the operation management of closed scenes, pedestrian trajectory prediction is one of the key technologies to realize intelligent decision-making. For example, a shopping mall needs to optimize shelf layout by predicting passenger flow trajectory, and a community security needs to identify abnormal stay behavior by trajectory prediction. The current mainstream pedestrian trajectory prediction technology mainly includes traditional time series models and deep learning models.
[0003] Traditional time series models such as LSTM (Long Short-Term Memory) and Kalman filter are based only on the historical trajectory of pedestrians themselves for prediction, without fully considering the interaction between pedestrians and surrounding objects in closed scenes, such as pedestrians avoiding shelves and following fellow travelers, resulting in a large deviation between the predicted results and the actual trajectory. In densely populated closed scenes, the prediction error is generally more than 1.5m, which cannot meet the actual needs of precise passenger flow guidance.
[0004] GAN models in deep learning models have the ability to generate diverse samples and are gradually applied to the field of trajectory prediction. However, existing GAN-based trajectory prediction techniques have obvious defects: first, the model structure is complex, usually containing multiple layers of encoders, decoders and independent attention modules, with a parameter size of more than 10 million, requiring high-performance servers to support, and cannot be deployed on edge devices; second, it is prone to "mode collapse" problem, with a trajectory repetition rate of more than 60%, making it difficult to cover the diversified behaviors of pedestrians such as temporary turning and avoiding; third, the training cycle is long, usually requiring hundreds of iterations, which is difficult to quickly adapt to scene changes in practical applications.
[0005] In summary, the existing technology lacks a pedestrian trajectory prediction method that takes into account "interaction relationship modeling, lightweight deployment, and low trajectory repetition rate", which cannot meet the practical needs of edge device deployment in closed scenes, and a targeted technical solution is urgently needed to solve the above problems. SUMMARY
[0006] To achieve the above purpose, the present application discloses a lightweight interactive perception GAN pedestrian trajectory prediction method, the core technical solution is as follows, and each step is connected in turn to form a complete technical link:
[0007] Step S101, based on the camera collected closed scene pedestrian historical trajectory data acquisition and preprocessing step: through the fixed installation in the key position of closed scene USB camera collects the historical trajectory data of target pedestrian, the collection frequency is set to 1Hz, ensure the real-time of data at the same time reduce the storage pressure;The historical trajectory data is stored in the form of“time stamp-two-dimensional coordinate(x, y)”key value pair, wherein the two-dimensional coordinate is calibrated based on the scene coordinate system(taking the lower left corner of the camera view as the origin, the horizontal direction as the x axis and the vertical direction as the y axis).In the preprocessing stage, the speed threshold method is used to eliminate abnormal data: calculate the difference value of two-dimensional coordinates of adjacent time stamps, combine the time interval(1s) to get the instantaneous speed, eliminate the abnormal trajectory points with speed>3m / s(out of the normal walking speed range of human), finally retain the effective historical trajectory data within the last 20 seconds(containing 20 trajectory points), provide high-quality data input for subsequent feature extraction.
[0008] Step S102, pedestrian and interactive object feature extraction step based on distance weight: first, determine the coordinate range of fixed interactive objects(shelves, railings, walls, etc.) in the closed scene through scene pre-calibration, identify the real-time coordinates of mobile interactive objects(other pedestrians) through target detection algorithm(such as YOLOv5 lightweight version);Then calculate the straight-line distance between the target pedestrian and each interactive object based on the Euclidean distance formula;According to the strength of the distance on the interaction, set the weight rule: when the straight-line distance≤2m(near distance strong interaction), the weight is set to 0.8, when the straight-line distance>2m and≤5m(middle distance weak interaction), the weight is set to 0.2, when the straight-line distance>5m(far distance no interaction), the weight is set to 0;Finally, arrange the“distance-weight”pairs of each interactive object in a fixed order to generate an interactive feature vector with a dimension of N×2(N is the number of interactive objects), realize the quantitative expression of the interaction relationship.
[0009] Step S103, a lightweight GAN model based on GRU+double constraint discriminator is constructed and trained: a lightweight GAN model is constructed, and the total parameters of the model are controlled within 800,000 to adapt to edge devices. The interactive perception generator adopts a series structure of "GRU time series encoder+single hidden layer interactive feature fusion layer": the GRU time series encoder contains 1 input layer (the input dimension is 20*2, corresponding to the two-dimensional coordinates of 20 track points), 1 GRU layer (the number of hidden units is 64), and 1 fully connected layer (the output dimension is 64, that is, the track feature vector); the interactive feature fusion layer fuses the track feature vector and the interactive feature vector (the fusion dimension is 64) in an element weighted sum manner. The double constraint discriminator adopts a lightweight fully connected structure, containing a distance constraint judgment unit and a diversity constraint judgment unit: the distance constraint judgment unit calculates the average Euclidean distance between the predicted trajectory and the real trajectory, and the diversity constraint judgment unit calculates the directional angle variance of multiple predicted trajectories. In the training stage, the cross-entropy loss function is used, and the effectiveness judgment result of the discriminator is used as the feedback signal, and the model is iteratively trained for 20-30 rounds. After each round of training, the proportion of effective predicted trajectories is calculated, and when the proportion is stable at more than 90%, the training is stopped, and the trained lightweight GAN model is obtained.
[0010] Step S104, trajectory prediction and effectiveness judgment based on the trained GAN model: input the effective historical trajectory data preprocessed in S101 into the trained interactive perception generator, extract the track time series features through the GRU time series encoder, fuse the interactive information through the interactive feature fusion layer, and then generate the predicted trajectory data through the generator output layer (fully connected layer, output dimension is 10*2, corresponding to the two-dimensional coordinates of the future 10 seconds). Then the double constraint discriminator performs effectiveness verification: the distance constraint judgment unit calculates the average Euclidean distance between the predicted trajectory and the real trajectory collected by the camera subsequently, and if it is <1m, it meets the physical rationality requirement; the diversity constraint judgment unit generates 10 predicted trajectories in succession, calculates the directional angle of the line connecting the starting point and the ending point of each trajectory, and if the directional angle variance is >15°, it meets the diversity requirement. Only when both constraints are met, the final predicted trajectory data is determined and output.
[0011] Step S105, historical and predicted trajectory differentiation visualization output step: the effective historical trajectory data in S101 and the final predicted trajectory data in S104 are transmitted to the display terminal synchronously, and visualization is realized through the Matplotlib library of Python: the historical trajectory is drawn with a gray solid line, and the predicted trajectory is drawn with a red dashed line. The starting point is marked as "Start", and the ending point is marked as "End". The prediction time and the average error are also displayed in the graph. The display terminal communicates with the edge computing module running the GAN model through Ethernet, ensuring that the data transmission delay is <0.1s. BRIEF DESCRIPTION OF DRAWINGS
[0012] Some specific embodiments of the present application will be described in detail below with reference to the accompanying drawings, which are presented by way of illustration and not of limitation. The flow of steps and main algorithms of the method are introduced in the form of pseudo code in the drawings. Those skilled in the art should understand that these drawings are not necessarily directly executable. The objects and features of the present application will be more apparent in view of the description below in conjunction with the drawings, in which:
[0013] Figure 1 A visual algorithm flowchart of the pedestrian trajectory prediction method of the lightweight interactive perception GAN according to the embodiment of the present application. DETAILED DESCRIPTION
[0014] In order to make the present application more apparent and understandable for its inventive points, the present application will be further described below in conjunction with the drawings and examples. In the following description, many details and specific examples are set forth in order to provide a more thorough understanding of the present application and to enable the present application to be fully conveyed to those skilled in the art. Although the present application can be implemented in many different ways other than the description, those skilled in the art can make corresponding extensions without departing from the scope and spirit of the present application, and therefore the present application is not limited by the specific examples and specific drawings disclosed below.
[0015] Although the present application has been described with reference to specific illustrative embodiments, it is not limited by these embodiments but only by the appended claims. Those skilled in the art should understand that the embodiments of the present application can be modified and changed without departing from the scope and spirit of the present application.
[0016] Figure 1 A flowchart of a pedestrian trajectory prediction method of a lightweight interactive perception GAN of an embodiment of the present application is shown. It includes the following steps: step S101, closed scene pedestrian historical trajectory data acquisition and preprocessing based on camera acquisition, step S102, pedestrian and interactive object feature extraction based on distance weight, quantitative expression of the current interactive relationship, step S103, lightweight GAN model construction and training based on GRU+double constraint discriminator, to obtain the trained lightweight GAN model, step S104, trajectory prediction and effectiveness judgment based on the trained GAN model, to obtain the final predicted trajectory data and output, step S105, historical and predicted trajectory differentiation visualization output. It can be widely applied to pedestrian behavior analysis scenes in closed scenes such as shopping mall passenger flow guidance, community security monitoring, supermarket dynamic line planning, etc., to provide data support for scene operation management and safety prevention and control.
[0017] While the application has been described with reference to particular illustrative embodiments, it is not intended to be limited to those embodiments but only to the appended claims. Those skilled in the art will appreciate that modifications can be made to the embodiments described without departing from the scope and spirit of the application.
Claims
1. The application discloses a lightweight interactive perception GAN pedestrian trajectory prediction method, the core technical solution is as follows, and each step is sequentially connected to form a complete technical link: Step S101, based on the camera collected closed scene pedestrian historical trajectory data acquisition and preprocessing step; Step S102, pedestrian and interactive object feature extraction step based on distance weight; Step S103, lightweight GAN model construction and training step based on GRU+double constraint discriminator; Step S104, trajectory prediction and effectiveness judgment step based on the trained GAN model; Step S105, historical and predicted trajectory differentiation visualization output step.
2. The pedestrian trajectory prediction method of claim 1, wherein The specific method of the camera collected closed scene pedestrian historical trajectory data acquisition and preprocessing step in step S101 is as follows: the historical trajectory data of the target pedestrian is collected through the camera fixedly installed at the key position of the closed scene, a reasonable collection frequency is set, the real-time data is ensured, and the storage pressure is reduced; the historical trajectory data is stored in the form of "time stamp-two-dimensional coordinate (x, y)" key value pair, wherein the two-dimensional coordinate is calibrated based on the scene coordinate system (taking the lower left corner of the camera view as the origin, the horizontal direction as the x axis, and the vertical direction as the y axis); In the preprocessing stage, the speed threshold method is used to eliminate abnormal data: the difference value of the two-dimensional coordinates of adjacent time stamps is calculated, the instantaneous speed is obtained combined with the time interval, and finally the effective historical trajectory data in the recent period is reserved, thereby providing high-quality data input for subsequent feature extraction.
3. The method of claim 1, wherein The specific method of the pedestrian and interactive object feature extraction step based on distance weight in step S102 is as follows: first, the coordinate range of the fixed interactive objects (shelves, railings, walls, etc.) in the closed scene is determined through scene pre-calibration, and the real-time coordinates of the moving interactive objects (other pedestrians) are recognized through a target detection algorithm (such as YOLOv5 lightweight version); then, the straight-line distance between the target pedestrian and each interactive object is calculated based on the Euclidean distance formula; according to the strength of the distance on the interaction, the corresponding weight rule is set: a larger weight is set for close-range strong interaction, a smaller weight is set for medium-distance weak interaction, and zero weight or extremely small weight is set for long-distance no interaction; finally, the "distance-weight" pairs of each interactive object are arranged in a fixed order to generate an interactive feature vector, and the quantitative expression of the interaction relationship is realized.
4. The pedestrian trajectory prediction method of claim 1, wherein The specific method of the lightweight GAN model construction and training step based on GRU+double constraint discriminator in step S103 is as follows: a lightweight GAN model is constructed, and the total parameters of the model are controlled so that it can adapt to edge devices; wherein the interactive perception generator adopts a series structure of "GRU time sequence encoder+single hidden layer interactive feature fusion layer": the GRU time sequence encoder contains 1 input layer, 1 GRU layer and 1 fully connected layer (output trajectory feature vector); The interaction feature fusion layer fuses the trajectory feature vector and the interaction feature vector by element weighted summation; the double-constraint discriminator adopts a lightweight fully connected structure, including a distance constraint judgment unit and a diversity constraint judgment unit; the distance constraint judgment unit calculates the average Euclidean distance between the predicted trajectory and the real trajectory, and the diversity constraint judgment unit calculates the direction angle variance of multiple predicted trajectories; In the training stage, the cross-entropy loss function is used, the effectiveness judgment result of the discriminator is used as the feedback signal, and the model is iteratively trained for several rounds. After each round of training, the proportion of effective predicted trajectories is calculated. When the proportion is stable and available, the training is stopped, and the trained lightweight GAN model is obtained.
5. The method of claim 1, wherein The specific method of the trajectory prediction and effectiveness judgment step of the trained GAN model in step S104 is as follows: input the preprocessed effective historical trajectory data in step S101 into the trained interaction perception generator, extract the trajectory time sequence features through GRU time sequence coding, fuse the interaction information through the interaction feature fusion layer, and then generate the predicted trajectory data through the generator output layer; Then, the double-constraint discriminator performs effectiveness verification: the distance constraint judgment unit calculates the average Euclidean distance between the predicted trajectory and the real trajectory collected by the camera subsequently, and the distance needs to meet the physical rationality requirement; the diversity constraint judgment unit generates several predicted trajectories in succession, calculates the direction angle of the line connecting the starting point and the ending point of each trajectory, calculates the direction angle variance, and the variance needs to meet the diversity requirement; only when both constraints are met, the final predicted trajectory data is determined and output.
6. The method of claim 1, wherein The specific method of the historical and predicted trajectory difference visualization output step in step S105 is as follows: the effective historical trajectory data in step S101 and the final predicted trajectory data in step S104 are transmitted to the display terminal simultaneously, and visualization is realized: the historical trajectory is drawn with a gray solid line, the predicted trajectory is drawn with a red dashed line, the starting point is marked as "Start", the ending point is marked as "End", and the prediction time and the average error are displayed on the graph.