A method and system for predicting pedestrian trajectories
By using a gated residual fusion mechanism and learnable pattern condition embedding, social interaction information is adaptively adjusted to generate stable candidate branches. This solves the problems of insufficient social interaction influence and unstable pattern features in pedestrian trajectory prediction, and improves the accuracy and consistency of trajectory prediction.
Patent Information
- Application Number
- CN202611455729.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-10
AI Technical Summary
Existing pedestrian trajectory prediction methods lack adaptive adjustment of the intensity of social interaction influence, lack stable pattern features in multimodal candidate generation, and fail to fully utilize pattern-related features across stages, resulting in insufficient modeling of intent and patterns.
By adaptively adjusting social interaction information through a gated residual fusion mechanism, using learnable pattern conditional embedding to drive the differentiated generation of multiple candidate branches, constructing a priori feature sequence of future motion, and combining it with a spatiotemporal representation network for decoding, a complete future trajectory is generated.
It improves the adaptability of endpoint intent modeling, enhances the stability of pattern features and the distinguishability of candidate endpoints, reduces pattern information loss, and improves the accuracy and consistency of trajectory prediction.
Smart Images

Figure CN123071979A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a prediction method in the field of artificial intelligence technology, particularly to a pedestrian trajectory prediction method and a pedestrian trajectory prediction system. Background Technology
[0002] Pedestrian trajectory prediction is a key supporting technology for intelligent agent safety decision-making in fields such as autonomous driving, intelligent transportation, and public safety. Its task is to predict the future movement path of pedestrians based on their historical trajectories. A pedestrian's future movement is influenced by both their own movement patterns and interactions with surrounding pedestrians, and may produce different movement outcomes even under the same historical conditions, thus exhibiting multimodal characteristics. With the development of deep learning technology, related research mainly focuses on social interaction modeling, multimodal prediction, and endpoint-guided trajectory generation. In terms of social interaction modeling, existing methods use social pooling, spatiotemporal graph convolution, or graph attention to aggregate features of neighboring pedestrians. In terms of multimodal prediction, existing methods use generative adversarial networks (GANs) and conditional variational autoencoders to generate multiple candidate trajectories. In terms of endpoint-guided trajectory generation, existing methods process endpoint prediction and complete future trajectory generation in stages, or gradually enhance the modeling ability for short-term dynamics and long-term dependencies through progressive pre-task learning.
[0003] Despite the progress made by the aforementioned methods in pedestrian trajectory prediction, the following technical problems remain. First, the intensity of social interaction influence lacks adaptive adjustment and insufficient utilization of dynamic motion states: some existing methods struggle to adjust the intensity of social interaction information based on actual interaction scenarios. Methods relying on static spatial proximity relationships fail to adequately utilize dynamic motion states such as relative speed, making it difficult to distinguish the differences in neighbor influence under different relative motion states. Second, multimodal candidate generation lacks stable pattern features and candidate differentiation mechanisms: some existing multimodal methods generate candidate trajectories through latent variable sampling or noise perturbation. Candidate differences are largely influenced by the random sampling process, making it difficult to form stable pattern features and achieve effective differentiation of different candidate trajectories. Third, insufficient utilization of pattern-related intermediate features across stages and insufficient adjustment of conditional effects at different future locations: some existing two-stage methods mainly transmit result-level conditions such as candidate endpoint coordinates, failing to fully utilize pattern-related intermediate features formed during the endpoint prediction stage; methods employing unified or relatively fixed fusion methods struggle to adjust the relative strength of historical motion information and future intention-related conditions according to future prediction locations, potentially affecting the consistency between short-term predictions and historical motion trends, and between long-term predictions and future intention-related conditions. Summary of the Invention
[0004] To address the shortcomings of existing pedestrian trajectory prediction methods in terms of intent and pattern modeling, this invention provides a pedestrian trajectory prediction method and system.
[0005] This invention is achieved using the following technical solution: a pedestrian trajectory prediction method, comprising the following steps: Temporal modeling was performed on the historical observation trajectories of each pedestrian in the same scene to obtain the corresponding temporal features; The local neighbors of the target pedestrian are determined at each observation time. The relative interaction state is constructed based on the relative position and velocity between the target pedestrian and each local neighbor. The soft interaction weight is calculated, and the local neighbor features are weighted and summed using the soft interaction weight to obtain the neighborhood aggregation feature. By combining the target pedestrian's own features and the difference between them and the neighborhood aggregation features, a social interaction correction feature is generated and cross-fused with the temporal features of the target pedestrian to obtain the social interaction feature; The individual intent and social interaction intent of the target pedestrian are formed by combining the temporal characteristics and social interaction characteristics to obtain the destination intent representation; The endpoint intent representation is fused with multiple learnable pattern conditions to obtain multiple pattern features, and the corresponding candidate endpoints are predicted to form multiple candidate branches. For each candidate branch, based on the historical conditions formed by the observed trajectory, the pattern features of the current candidate branch are used as pattern conditions, and the historical conditions and pattern conditions are weighted by the fusion weights related to the future predicted position to construct a sequence of prior features of future motion. The observed trajectory feature sequence, the future motion prior feature sequence, and the candidate endpoint features of the current candidate branch are combined into an input sequence, which is then input into a spatiotemporal representation network for decoding to generate a complete future trajectory.
[0006] This invention adaptively adjusts the correction strength of social interaction information on individual intentions through a gated residual fusion mechanism, drives the differentiated generation of multiple candidate branches using learnable pattern conditional embedding, and constructs a priori feature sequence of future motion using fusion weights related to future predicted positions. This solves the technical problems of insufficient intention and pattern modeling in existing pedestrian trajectory prediction methods, which are caused by the lack of adaptive adjustment of the influence strength of social interaction, the instability of multimodal candidate generation pattern features, and the insufficient utilization of pattern-related features across stages.
[0007] As a further improvement to the above scheme, the method for obtaining the neighborhood aggregation feature specifically includes the following steps: At each observation time, a preset number of nearest neighbors are selected for the target pedestrian based on the Euclidean distance between their absolute coordinates, forming a local nearest neighbor set. The relative position vector and relative velocity vector between the target pedestrian and each local nearest neighbor are concatenated to obtain the relative interaction state vector; The relative interaction state vector is input into the weight network and softmax normalized within the local nearest neighbor set to obtain the soft interaction weight. The neighborhood aggregation feature is obtained by weighting and summing the features of each neighbor in the local nearest neighbor set using the soft interaction weight.
[0008] As a further improvement to the above scheme, the target pedestrian's own features are concatenated with the difference obtained by subtracting the neighborhood aggregation features from the target pedestrian's own features, and the concatenation result is input into the correction network to obtain the social interaction correction features; the social interaction features are obtained by calculating cross-attention using the temporal features of the target pedestrian as the query and the social interaction correction feature sequence as the key and value.
[0009] As a further improvement to the above scheme, the method for obtaining the endpoint intent representation specifically includes the following steps: Attention is converged on the temporal features and social interaction features of the target pedestrian to obtain the individual intent vector and the social interaction intent vector. The individual intent vector and the social interaction intent vector are concatenated and then gating vectors are generated using the Sigmoid function. The social interaction intention vector is gated using the gating vector, and the gating result scaled by a preset weight coefficient is added to the individual intention vector to obtain the endpoint intention representation.
[0010] As a further improvement to the above scheme, the endpoint intention representation is concatenated with each mode condition embedding vector and then input into a shared mode mapping network to obtain multiple mode feature vectors; each mode feature vector is then input into a shared coordinate prediction network to obtain the corresponding candidate endpoint coordinates.
[0011] As a further improvement to the above scheme, the method for constructing the future motion prior feature sequence specifically includes the following steps: For the first time period excluding the end time within the prediction period For each future moment, obtain the learnable basic feature vector and temporal position embedding vector at that moment; Construct the motion prior feature vector for this future moment using the following formula: in, For the first Learnable basic features corresponding to each future moment For the first Embedding of the temporal position of a future moment, The historical conditions formed by the temporal characteristics and social interaction characteristics of the target pedestrian. This represents the pattern feature vector of the current candidate branch; and The fusion weights, which are related to the future predicted location, are used to adjust the influence of historical conditions and model conditions on the first... The strength of the influence of prior features of future motion.
[0012] As a further improvement to the above scheme, the calculation formula for the social interaction correction feature is as follows: In the formula, For the social interaction correction features, To calibrate the network; The characteristics of the target pedestrian; This is a neighborhood aggregation feature.
[0013] As a further improvement to the above scheme, the calculation formula for the endpoint intention representation is as follows: In the formula, The intended endpoint is represented as follows. For the individual's intention, For preset weighting coefficients, Let be the gate vector. This indicates element-wise multiplication. The stated social interaction intention; The formula for calculating the candidate endpoint is: in, For the pattern features, For pattern mapping networks, For the first Conditional embedding of patterns, The candidate endpoint. For coordinate prediction networks, This is a candidate branch.
[0014] As a further improvement to the above scheme, the network parameters of the prediction method are obtained through a phased training process, and the phased training process includes: In the short-term prediction training phase, it uses real trajectories as input to train the spatiotemporal representation network to establish the ability to model time series. In the endpoint prediction training phase, the intention fusion and pattern branch generation parts are trained on the basis of the spatiotemporal representation network, and differential constraints that suppress pattern feature convergence and suppress excessive concentration of candidate endpoints are introduced. In the complete trajectory prediction training phase, the training endpoint prediction branch and trajectory generation branch are combined.
[0015] The present invention also provides a pedestrian trajectory prediction system, the system being used to execute any of the above-described pedestrian trajectory prediction methods, the system comprising: The intent generation module performs the following processes: It performs temporal modeling on the historical observation trajectories of each pedestrian in the same scene to obtain the temporal features of each pedestrian; it determines the local nearest neighbors of the target pedestrian at each observation time, constructs a relative interaction state based on the relative position and relative speed between the target pedestrian and each local nearest neighbor, calculates soft interaction weights, and uses the soft interaction weights to perform a weighted summation of the local nearest neighbor features to obtain neighborhood aggregation features; it generates social interaction correction features by combining the target pedestrian's own features and the difference between them and the neighborhood aggregation features, and cross-fuses the social interaction correction features with the temporal features of the target pedestrian to obtain social interaction features; it forms individual intent and social interaction intent from the temporal features of the target pedestrian and the social interaction features, and obtains the endpoint intent representation through gated residual fusion. The pattern branch generation module is used to embed and fuse the endpoint intent representation with multiple learnable pattern conditions to obtain multiple pattern features, and predict the corresponding candidate endpoints by each pattern feature to form a one-to-one "pattern feature - candidate endpoint" candidate branch. The trajectory decoding module performs the following processing on each candidate branch: based on the historical conditions of the observed trajectory formation, and using the pattern features of the current candidate branch as pattern conditions; weighting the historical conditions and the pattern conditions using fusion weights related to the future predicted location to construct a future motion prior feature sequence; combining the observed trajectory feature sequence, the future motion prior feature sequence, and the candidate endpoint features of the current candidate branch into an input sequence, and inputting the input sequence into a spatiotemporal representation network for decoding to generate a complete future trajectory.
[0016] Compared with existing pedestrian trajectory prediction methods, the pedestrian trajectory prediction method and system of the present invention have the following advantages: 1. This pedestrian trajectory prediction method adaptively adjusts the correction strength of social interaction information on individual intentions through a gated residual fusion mechanism, drives the differentiated generation of multiple candidate branches using learnable pattern conditional embedding, and constructs a prior feature sequence of future motion using fusion weights related to the future predicted position. This solves the technical problems of insufficient intention and pattern modeling caused by the lack of adaptive adjustment of the influence of social interaction, the instability of multimodal candidate generation pattern features, and the insufficient utilization of pattern-related features across stages in existing pedestrian trajectory prediction methods.
[0017] 2. This pedestrian trajectory prediction method improves the adaptability of endpoint intent modeling under different interaction scenarios. The method dynamically determines local neighbors at each historical observation time, calculates soft interaction weights using relative position and relative velocity, and adaptively adjusts the correction strength of social interaction information for individual intent through gated residual fusion. Compared with methods that use fixed neighborhood relationships or fuse social interaction information at a fixed ratio, this method can flexibly adjust the degree of intervention of social interaction influence according to the actual interaction scenario, reduce the bias in judging interaction relationships, and make the endpoint intent representation more accurately reflect the real movement trend of pedestrians under different interaction scenarios.
[0018] 3. This pedestrian trajectory prediction method improves the stability of pattern features and the distinguishability of candidate endpoints. The method uses multiple learnable pattern conditional embeddings to drive the generation of corresponding pattern features and candidate endpoints, ensuring a one-to-one correspondence between the pattern conditional embeddings, pattern features, and candidate endpoints, and avoiding random sampling or noise perturbation as the main source of candidate differences. It suppresses pattern feature convergence through pattern feature difference constraints and suppresses excessive concentration of candidate endpoints through candidate endpoint difference constraints, enabling each candidate branch to possess stable pattern representation and clear motion tendency differentiation, thereby improving the candidate set's coverage of various future motion possibilities.
[0019] 4. This pedestrian trajectory prediction method improves the ability to transfer and utilize pattern features between endpoint prediction and trajectory generation. The method transmits candidate endpoints and their corresponding pattern features as conditions to the trajectory generation stage. The same pattern feature participates in both candidate endpoint prediction and the construction of the prior feature sequence of future motion for the corresponding candidate branch, effectively reducing the loss of pattern information between endpoint prediction and trajectory generation. Simultaneously, by adjusting the relative strength of historical and pattern conditions through fusion weights related to the future predicted position, trajectory prediction closer to the observation end relies more on historical motion trends, while trajectory prediction closer to the prediction end focuses more on pattern conditions, thus balancing the continuity of future trajectory motion and pattern consistency.
[0020] 5. The beneficial effects of this pedestrian trajectory prediction system are the same as those of the pedestrian trajectory prediction methods described above, and will not be repeated here. Attached Figure Description
[0021] Figure 1 This is a flowchart of the pedestrian trajectory prediction method according to Embodiment 1 of the present invention.
[0022] Figure 2 This is a visualization of the trajectory prediction results of the pedestrian trajectory prediction method of Embodiment 3 of the present invention on the ETH / UCY dataset.
[0023] Figure 3This is a schematic diagram of the network structure and phased training of the prediction model in Embodiment 4 of the present invention. Part (a) shows the overall process of phased training, part (b) shows the internal structure of the spatiotemporal representation network, and part (c) shows the processing details of each training stage.
[0024] Figure 4 This is a diagram illustrating the overall technical solution of the pedestrian trajectory prediction system in Embodiment 5 of the present invention.
[0025] Figure 5 for Figure 4 A schematic diagram of the intent generation module in the pedestrian trajectory prediction system.
[0026] Figure 6 for Figure 4 A schematic diagram of the trajectory decoding module in the pedestrian trajectory prediction system. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0028] Example 1 Please see Figure 1 This embodiment provides a pedestrian trajectory prediction method, specifically a pedestrian trajectory prediction method based on destination intention and pattern condition modeling. This prediction method forms a complete processing chain from historical observed trajectories to multiple candidate future trajectories, following the order of intention generation, pattern branch generation, and trajectory decoding. The prediction method includes the following steps.
[0029] (1) Perform time-series modeling on the historical observation trajectories of each pedestrian in the same scene to obtain the time-series features of each pedestrian. This step primarily involves temporal modeling of the historical observation trajectory of the target pedestrian to obtain temporal features that characterize its own movement trend during the observation period. This provides a foundation for the subsequent construction of individual intentions.
[0030] (2) Determine the local nearest neighbors of the target pedestrian at each observation time. Construct a relative interaction state based on the relative position and relative speed between the target pedestrian and each local nearest neighbor. Calculate the soft interaction weights and use the soft interaction weights to perform a weighted summation of the local nearest neighbor features to obtain the neighborhood aggregation features. The method for obtaining the neighborhood aggregation features specifically includes the following steps: (2.1) At each observation time, select a preset number of nearest neighbors for the target pedestrian based on the Euclidean distance between the absolute coordinates to form a local nearest neighbor set; (2.2) Concatenate the relative position vector and relative speed vector between the target pedestrian and each local nearest neighbor to obtain a relative interaction state vector; (2.3) Input the relative interaction state vector into the weight network and perform softmax normalization within the local nearest neighbor set to obtain the soft interaction weights; (2.4) Use the soft interaction weights to perform a weighted summation of the features of each nearest neighbor within the local nearest neighbor set to obtain the neighborhood aggregation features. The aforementioned steps determine the local neighbors of the target pedestrian based on the spatial distance between pedestrians, and construct a relative interaction state based on the relative position and relative speed of the target pedestrian and each local neighbor. The local neighbor relationship is dynamically updated with the relative position of the pedestrian, and the soft interaction weight is dynamically adjusted with the relative position and relative speed.
[0031] (3) Generate social interaction correction features by combining the characteristics of the target pedestrian and the difference between them and the aggregated features of the neighborhood, and combine the social interaction correction features with the temporal features of the target pedestrian. Cross-integration yields social interaction characteristics. In this embodiment, the target pedestrian's own features are concatenated with the difference obtained by subtracting the neighborhood aggregation features from the target pedestrian's own features, and the concatenation result is input into the correction network to obtain the social interaction correction features. This embodiment uses the temporal features of the target pedestrian. As a query, the social interaction correction feature sequence is used as the key and value, and the social interaction features are calculated through cross-attention. Social interaction characteristics Characterize the impact of surrounding pedestrian interactions on the target pedestrian's future destination, and use this as a basis for constructing social interaction intentions.
[0032] (4) Based on the temporal characteristics of the target pedestrian Social interaction characteristics Individual intentions and social interaction intentions are formed, and the endpoint intention representation is obtained through gating residual fusion. Here, the gating weights are calculated from both, and the gating-regulated social interaction intention is superimposed on the individual intention as a residual correction term to obtain the endpoint intention representation. This representation preserves the target pedestrian's own movement trend and incorporates social interaction information under gating regulation.
[0033] In this embodiment, the method for obtaining the destination intent representation specifically includes the following steps: (4.1) analyzing the temporal features of the target pedestrian. Social interaction characteristics Attention convergence is performed to obtain individual intention vector and social interaction intention vector; (4.2) The individual intention vector and social interaction intention vector are concatenated and then gating vector is generated by the Sigmoid function; (4.3) The gating vector is used to gating the social interaction intention vector, and the gating adjustment result scaled by the preset weight coefficient is added to the individual intention vector to obtain the endpoint intention representation.
[0034] (5) Represent the intended endpoint with multiple ( Multiple learnable pattern conditional embeddings are fused to obtain multiple pattern features, and each pattern feature predicts its corresponding candidate endpoint, forming multiple candidate branches. In this embodiment, the endpoint intention representation is concatenated with each pattern conditional embedding vector and then input into a shared pattern mapping network to obtain multiple pattern feature vectors. Furthermore, in this embodiment, each pattern feature vector is input into a shared coordinate prediction network to obtain the corresponding candidate endpoint coordinates.
[0035] The same endpoint is intended to represent different endpoints. Each pattern condition is embedded and fused to obtain... Each pattern feature predicts a corresponding candidate endpoint, forming... Each candidate branch corresponds to a "pattern feature - candidate endpoint". The pattern feature is also passed to the next step as a condition for generating a complete future trajectory for the corresponding candidate branch, so that the pattern-related features in the candidate endpoint prediction stage continue to the future trajectory generation stage, and constrain the same candidate branch in both stages.
[0036] (6) For each candidate branch, based on the historical conditions formed by the observed trajectory, the pattern features of the current candidate branch are used as the pattern conditions, and the historical conditions and pattern conditions are weighted by the fusion weights related to the future predicted position to construct the future motion prior feature sequence.
[0037] In this embodiment, the temporal features and social interaction features of the observed trajectory are aggregated and fused to form historical conditions, which represent the movement trend and social interaction status during the observation phase. The pattern features of the current candidate branch serve as pattern conditions, and together with the historical conditions, construct a sequence of prior features for future movement. The fusion weights related to the future prediction location adjust the relative strength of the two types of conditions along the future time step, so that the prior features closer to the observation end emphasize the historical movement trend, while the prior features closer to the prediction end emphasize the pattern conditions of the current candidate branch, thus balancing short-term movement continuity and long-term pattern consistency.
[0038] (7) Combine the observed trajectory feature sequence, the future motion prior feature sequence and the candidate endpoint feature into an input sequence (joint feature sequence), and input the input sequence into the spatiotemporal representation network for decoding to generate a complete future trajectory.
[0039] In some embodiments, the predicted relative bias is added to the candidate endpoint as a reference to obtain the position of the generated trajectory at the final time, thereby enhancing the consistency between the trajectory and the candidate endpoint. Each candidate branch executes this process to obtain... A candidate future trajectory.
[0040] In summary, compared with existing pedestrian trajectory prediction methods, the pedestrian trajectory prediction method in this embodiment has the following advantages: 1. This pedestrian trajectory prediction method adaptively adjusts the correction strength of social interaction information on individual intentions through a gated residual fusion mechanism, and drives the differentiated generation of multiple candidate branches by using learnable pattern conditional embedding. At the same time, it constructs a future motion prior feature sequence through fusion weights related to future prediction positions. This solves the technical problems of insufficient intention and pattern modeling caused by the lack of adaptive adjustment of the influence of social interaction, the instability of multimodal candidate generation pattern features, and the insufficient utilization of pattern-related features across stages in existing pedestrian trajectory prediction methods.
[0041] 2. This pedestrian trajectory prediction method improves the adaptability of endpoint intent modeling under different interaction scenarios. The method dynamically determines local neighbors at each historical observation time, calculates soft interaction weights using relative position and relative velocity, and adaptively adjusts the correction intensity of individual intent based on social interaction information through gated residual fusion. Compared with methods that use fixed neighborhood relationships or fuse social interaction information at a fixed ratio, this method can flexibly adjust the degree of intervention of social interaction influence according to the actual interaction scenario, reduce the bias in judging interaction relationships, and make the endpoint intent representation more accurately reflect the true movement trend of pedestrians under different interaction intensities.
[0042] 3. This pedestrian trajectory prediction method improves the stability of pattern features and the distinguishability of candidate endpoints. The method uses multiple learnable pattern conditional embeddings to drive the generation of corresponding pattern features and candidate endpoints, ensuring a one-to-one correspondence between the pattern conditional embeddings, pattern features, and candidate endpoints, and avoiding random sampling or noise perturbation as the main source of candidate differences. It suppresses pattern feature convergence through pattern feature difference constraints and prevents excessive concentration of candidate endpoints through candidate endpoint difference constraints, enabling each candidate branch to possess stable pattern representation and clear motion tendency differentiation, thereby improving the candidate set's coverage of various future motion possibilities.
[0043] 4. This pedestrian trajectory prediction method improves the ability to transfer and utilize pattern features between endpoint prediction and trajectory generation. The method transmits candidate endpoints and their corresponding pattern features as conditions to the trajectory generation stage. The same pattern feature participates in both candidate endpoint prediction and the construction of the prior feature sequence of future motion for the corresponding candidate branch, effectively reducing the loss of pattern information between endpoint prediction and trajectory generation. Simultaneously, by adjusting the relative strength of historical and pattern conditions through fusion weights related to future prediction positions, trajectory prediction closer to the observation end relies more on historical motion trends, while trajectory prediction closer to the prediction end focuses more on pattern conditions, thus balancing the continuity of future trajectory motion and pattern consistency.
[0044] Example 2 This embodiment provides a pedestrian trajectory prediction method, which adds a processing scenario to the method described in Embodiment 1. Within the processing scenario... The two-dimensional planar trajectory of a pedestrian. Let the length of the observation period be denoted as . The predicted time period length is The length of the entire trajectory is Given each pedestrian during the observation period This method predicts a historical location within a given timeframe for future periods. The two-dimensional position within. Let the number of pattern conditions be . , The value is 20. In this embodiment, Take 8, Taking 12 means using 8 historical trajectories to predict 12 future trajectories.
[0045] I. Data Representation and Preprocessing The prediction method in this embodiment takes as input the planar positions of all pedestrians in the scene at various times during the observation period. The pedestrian to be predicted is called the target pedestrian, and the remaining pedestrians constitute its potential interaction objects. (Pedestrians are recorded.) At any moment The position is Its velocity is obtained by the position difference between adjacent time points, that is: .
[0046] Trajectory prediction relies on the movement patterns of the target pedestrians themselves and the spatial interaction relationships between pedestrians. In this embodiment, relative coordinates and absolute coordinates are used to represent these two types of information. Relative coordinates (normalized coordinates) are obtained by translating the trajectory, with the position of the target pedestrian at the last observation time as the reference origin, and are used to learn movement patterns that are independent of the absolute position of the scene. Absolute coordinates are the original coordinates of each pedestrian in the scene coordinate system, and are used to determine local neighbors and calculate relative interaction states.
[0047] II. Spatiotemporal Representation Network The spatiotemporal representation network is the foundational network shared by the intent generation module and the trajectory decoding module. It consists of multiple stacked spatiotemporal representation units with identical structures, as shown in the figure. Each unit can operate along either the sequence modeling path or the feature extraction path. Both paths use the same components, differing only in information organization and output. The sequence modeling path processes the input sequence layer by layer, outputting a feature sequence of equal length to the input for decoding in the next layer or subsequent layers. The feature extraction path outputs temporal features characterizing the motion patterns of individuals from the observed sequence. And characterizing the social interaction features of mutual influence .
[0048] When spatiotemporal representation networks perform deep modeling of a single sequence, each layer adopts a sequence modeling path. For example, it processes joint sequences composed of observation features, prior features of future motion, and endpoint features; and extracts data from the observation sequence. and In this approach, the preceding layers use a sequence modeling path, while only the last layer uses a feature extraction path. The sequence modeling path does not incorporate social interaction information between pedestrians; social interaction modeling is only used in the feature extraction path.
[0049] The speed-aware graph network takes the feature sequences and absolute coordinate sequences of each pedestrian as input, performs selection, weighting and exclusion in sequence, and outputs social interaction correction features to generate social interaction features for feature extraction paths.
[0050] Select: Each observation time Based on the Euclidean distance between absolute coordinates, several nearest neighbors are determined for the target pedestrian i and a local nearest neighbor set is formed. The interaction edges between the target pedestrian and each local nearest neighbor constitute the local interaction graph at that moment. The nearest neighbor set is determined separately for each observation moment and is updated moment by moment as the relative position of the pedestrian changes, rather than being fixed all at once. In one implementation, four nearest neighbors are selected at each moment.
[0051] Weighted: For target pedestrians Its neighbors The relative position and relative velocity are calculated from absolute coordinates, forming a relative interactive state: Input the relative interaction states into a learnable weight network The soft interaction weights are then normalized within the nearest neighbor set. The soft interaction weights continuously and differentiably adjust the degree of influence of different neighbors within the local nearest neighbor set, avoiding the use of the same contribution coefficient for all selected neighbors. The weights depend on both relative position and relative velocity, which can distinguish the influence of neighbors under different relative motion states.
[0052] Exclusion: Nearest neighbor features are weighted and aggregated using soft interaction weights to obtain neighborhood aggregated features. Then, the characteristics of the target pedestrian. The difference between the target pedestrian's own features and the neighborhood aggregation features. The parts are spliced together and then processed by a learnable correction network. Generate social interaction correction features: In the formula, For social interaction correction features, To calibrate the network. Based on the characteristics of the target pedestrian, Indicates neighbor At any moment Its characteristics. This is a neighborhood aggregation feature. It explicitly characterizes the deviation of the target pedestrian's motion state relative to the neighborhood as a whole, allowing the correction features to focus on the corrections that the interaction produces for the pedestrian.
[0053] The sequence modeling path deepens the modeling of the input feature sequence layer by layer: Input feature sequence Time-series modeling network Obtaining temporal features The query is generated using time-series features, and the input features are linearly projected. Generate keys and values and perform attention calculations: Attention computation captures global dependencies within the sequence. Input features, temporal features, and attention output are fused in a residual manner and then passed through a feedforward network. The output is obtained by connecting the residuals: The three residual fusions preserve and superimpose the original input, temporal modeling results, and global dependency results at each layer, which is beneficial for information transfer and training stability when stacking deep layers. The attention key values of the sequence modeling path come from the sequence itself and do not introduce social interaction information between pedestrians. In one implementation, the temporal modeling network is implemented by a state-space model, and the feature extraction capability is enhanced by stacking several temporal representation units.
[0054] The feature extraction path extracts temporal features and social interaction features from the observed trajectory. The input feature sequence is processed by a temporal modeling network to obtain temporal features. Describe the movement sequence pattern of the target pedestrian; then use Generate queries, generate keys and values from the social interaction correction feature sequence U output by the speed-aware graph network, and perform cross-fusion to obtain social interaction features: The obtained social interaction characteristics Used to characterize the motion context after social interaction correction. The feature extraction path directly outputs temporal features. Social interaction characteristics This is used for intent generation and trajectory decoding, respectively.
[0055] III. Intent Generation Intent generation primarily uses the historical observation trajectories of the target pedestrian and other pedestrians in the same scene as input. Based on the target pedestrian's historical observation trajectory and its interaction information with local neighbors, it generates the target pedestrian's destination intent representation. Internally, it includes three parts: temporal feature extraction, social interaction feature generation, and gated intent fusion.
[0056] Temporal feature extraction. Trajectory encoder. The normalized coordinates of the target pedestrian at each observation time are encoded into an observation marker sequence, which is then superimposed on the embedded data to identify the type of observation marker and input into a stacked spatiotemporal representation network. The preceding layers adopt a sequence modeling path, and the final layer, a temporal modeling network, outputs the temporal features of the observation period. .
[0057] Social interaction feature generation. The velocity-aware graph network sequentially performs three steps—selection, weighting, and rejection—to generate a social interaction correction feature sequence U. The final layer uses... Generate queries, using U to generate keys and values, and obtain social interaction features through cross-fusion. .
[0058] Gated residuals are intended to be fused. and Learnable attention convergence is performed separately for feature sequences; The scores were assigned based on the importance of each moment, normalized, and then weighted and summed. In the formula, This is a learnable scoring vector. Attention convergence yields individual intent. Social interaction intentions Jointly determine the gating vector: in The sigmoid function is used. The gating vector adjusts the correction strength of the social interaction intention dimension by dimension. The adjusted social interaction intention is then superimposed onto the individual intention as a residual to obtain the endpoint intention representation. In the formula, The intention to indicate the endpoint For individual intentions, For preset weighting coefficients, For the gated vector, This indicates element-wise multiplication. For social interaction intentions.
[0059] IV. Pattern Branch Generation When generating pattern branches, pre-set Learnable pattern conditional embedding Each pattern condition embedding corresponds to a potential future motion pattern. To express the intention of the destination Each pattern is embedded and fused with its own modal conditions, and then processed through a shared pattern mapping network. get Each pattern feature is then processed by a shared coordinate prediction network. Generate corresponding candidate endpoints: in, For pattern features, For pattern mapping networks, For the first Conditional embedding of patterns, As a candidate endpoint, For coordinate prediction networks, This is a candidate branch.
[0060] Each candidate endpoint is derived from the same endpoint intent under different modal conditions. Each candidate branch shares a unified intent, but differentiates itself through various pattern conditions, thus structurally covering multiple possibilities for future movements. Pattern characteristics As a mode condition, it is carried over to the trajectory decoding module, so that the candidate endpoints of the same candidate branch and the subsequent trajectory generation are constrained by the same mode conditions, which is conducive to maintaining consistency within the branch.
[0061] V. Trajectory Decoding The spatiotemporal representation network of the trajectory generation branch processes the observed trajectory according to the feature extraction path to obtain temporal features and social interaction features; the two types of features are then fused after attention convergence to obtain historical conditions. The method for constructing the prior feature sequence of future motion specifically includes the following steps: For the first time period excluding the end time within the prediction period For each future moment, first obtain the learnable basic feature vector and temporal position embedding vector for that moment, and then construct the motion prior feature vector for that future moment using the following formula: in, For the first Learnable basic features corresponding to each future moment For the first Embedding of the temporal position of a future moment, The historical conditions formed by the temporal characteristics and social interaction characteristics of the target pedestrian. This is the pattern feature vector of the current candidate branch. and The fusion weights, which are related to the future predicted location, are used to adjust the influence of historical conditions and model conditions on the first... The strength of the influence of prior features of future motion.
[0062] This embodiment predicts all future times within the prediction period, excluding the final time. Prior features for future motion are constructed separately to form a sequence of prior features for future motion. Different future positions are assigned corresponding fusion weights, ensuring that historical conditions and pattern conditions have different relative effects at different future positions. To establish a basic weight distribution that gradually transitions from historical motion trends to candidate pattern conditions, this embodiment sets a baseline weight that increases with future position. Then, position-by-position learnable bounded offsets are superimposed on these baseline weights to obtain the final mode conditional weights. : in To establish a baseline weight that increases with future position, a basic trend is gradually transitioned from historical conditions to pattern conditions. For the first The learnable offset for a future position is subject to a bounded transformation. and These represent the lower and upper bounds of the model condition weights, respectively. The baseline weights ensure that future positions closer to the observation end retain more historical motion information, while future positions closer to the prediction end reflect more model conditions. The model can also adjust the fusion strength of each future position based on training data.
[0063] Observation trajectory feature sequence The prior feature sequence of future motion and the endpoint features obtained by the trajectory encoder from the candidate endpoint of the candidate branch are concatenated in time to form a joint sequence: A joint sequence overlay type embedding is used to distinguish between three types of features: observation, future, and endpoint. After the spatiotemporal representation network of the input trajectory generation branch is modeled layer by layer, the decoder outputs the position ŷ_τ at each moment within the prediction period. Observation features provide historical facts, prior features provide time-by-time generation guidance, and endpoint features provide the target constraint for the current candidate branch. The three types of features work together to ensure that the generated trajectory continues the historical motion and is guided by the candidate endpoint. In one implementation, the decoded features of the endpoint are used to predict the bias relative to the candidate endpoint. The position of the endpoint is obtained by adding the bias to the candidate endpoint, i.e.: right Each candidate branch performs the above process to obtain... A complete candidate future trajectory.
[0064] VI. Reasoning Process The reasoning phase does not use future information: the endpoint prediction branch obtains the endpoint intent representation in one step. Each pattern feature and One candidate endpoint; parallel processing of trajectory generation branches. Each of the candidate branches constructs a priori information for future motion and generates a complete trajectory, outputting... Structured candidate future trajectories.
[0065] Example 3 This embodiment provides a pedestrian trajectory prediction method, which is experimentally verified based on Embodiment 1 or 2. This embodiment constructs an experimental model named SoMI-Mamba and conducts experiments on the Stanford Unmanned Aerial Vehicle Dataset (SDD) and the ETH / UCY dataset, respectively. The experiment uses 8 historical observation locations to predict 12 future locations and generates 20 candidate future trajectories for each target pedestrian.
[0066] Please see Figure 2 This chart displays the trajectory prediction results for five scenarios in the ETH / UCY dataset. The five columns correspond to the ETH, HOTEL, UNIV, ZARA1, and ZARA2 scenarios, respectively. The top image for each scenario shows the 20 candidate future trajectories generated by the model, while the bottom image shows the trajectory with the smallest average displacement error selected from these candidate trajectories, using the actual future trajectory as a reference.
[0067] Quantitative evaluation uses minimum average displacement error and minimum final displacement error , representing the minimum average displacement error and the minimum final displacement error among the 20 candidate trajectories, respectively. The candidate trajectories corresponding to these two values can be different. The smaller both metrics are, the more accurate the prediction result. The SDD dataset uses pixels as the unit, while the ETH / UCY dataset uses meters.
[0068] Table 1 shows the trajectory prediction errors of the method of this invention compared to existing methods on the SDD dataset. The method of this embodiment... and The resolutions are 6.86 pixels and 10.33 pixels, respectively, which are approximately 2.4% and 3.0% lower than the best comparison results for the corresponding indicators in the table.
[0069] Table 1. Trajectory prediction error table on the SDD dataset Table 2 shows the trajectory prediction errors of the method in this embodiment compared to existing methods on the ETH / UCY dataset. The average error of the method in this invention across five scenarios is... and average The measurements were 0.19 meters and 0.30 meters respectively, both lower than the corresponding average values of the existing methods in the table.
[0070] Table 2. Trajectory prediction error on the ETH / UCY dataset Example 4 Please see Figure 3 This embodiment provides a pedestrian trajectory prediction method. Based on Embodiment 1 or 2, it obtains a corresponding prediction model by training the network parameters of the prediction method in stages, thereby sequentially establishing the ability to perform time-series modeling, the ability to predict the destination, and the ability to generate a complete trajectory. The staged training process includes the following stages.
[0071] (a) In the short-term prediction training phase, the spatiotemporal representation network is trained using real trajectories as input to establish temporal modeling capabilities. The input is the previous real trajectory. Frame, i.e., frame 1 to frame 2 All frames are represented using normalized coordinates. The spatiotemporal representation network processes data layer by layer according to the sequence modeling path, using only the current and previous trajectory information at each time step. The prediction head predicts the position of the next time step based on the output features of each time step, predicting the second frame from the first frame, the third frame from the first and second frames, and so on, frame by frame. The training loss is the time-by-time prediction error: in The actual location is used as input. This stage uses the actual trajectory as input, allowing the spatiotemporal representation network to focus on building basic temporal modeling capabilities.
[0072] (b) In the endpoint prediction training phase, the intent fusion and pattern branch generation parts are trained based on the spatiotemporal representation network, and differential constraints are introduced to suppress pattern feature convergence and suppress excessive concentration of candidate endpoints. The training loss balances accuracy and diversity. The first term is the minimum distance reconstruction loss, which only calculates the error of the candidate endpoint that is closest to the true endpoint, driving the optimal candidate to approach the true endpoint; the latter two terms are used to enhance the candidate difference, which respectively penalize the cosine similarity between each pair of pattern features that exceeds the preset margin and the spatial concentration between each candidate endpoint, suppressing pattern collapse convergence and causing the candidate endpoints to be spatially dispersed. and The preset weight coefficients are used. In one implementation, this stage includes warm-up and joint fine-tuning: during warm-up, the spatiotemporal representation network obtained in stage one is frozen, and only the newly added gated intent fusion, pattern branch generation, and endpoint prediction heads are trained; during joint fine-tuning, the freeze is lifted and all parameters are trained.
[0073] (c) In the complete trajectory prediction training phase, the endpoint prediction branch and the trajectory generation branch are jointly trained. During training, the endpoint prediction branch is used to obtain... After considering 10 candidate endpoints, select the branch whose candidate endpoint is closest to the true endpoint: The candidate endpoint of this branch is denoted as the optimal candidate endpoint. (Right now The optimal candidate endpoint and its corresponding pattern features are fed into the trajectory generation branch to generate a complete future trajectory and apply supervision. The training loss is mainly based on trajectory regression error, supplemented by endpoint supervision. in, For the future The true location The weighting coefficients are preset. In one implementation, the trajectory generation branch is trained with a main learning rate, while the endpoint prediction branch is fine-tuned with a smaller learning rate to protect the converged endpoint prediction capability; the optimal candidate endpoint and pattern features stop gradient backpropagation before being fed into the trajectory generation branch to avoid the trajectory-side gradient affecting the endpoint prediction branch through the discrete optimal selection path.
[0074] Example 5 Please see Figure 4 , Figure 5 as well as Figure 6This embodiment provides a pedestrian trajectory prediction system, which is used to execute the pedestrian trajectory prediction method provided in any of the embodiments 1-4. The prediction system includes an intent generation module, a pattern branch generation module, and a trajectory decoding module. The intent generation module processes historical observed trajectories to form a destination intent representation that integrates individual movement trends and social interaction information. The pattern branch generation module expands the destination intent representation into multiple pattern features and their corresponding candidate destinations, forming multiple candidate branches. The trajectory decoding module generates the complete future trajectory corresponding to each candidate branch. These three modules are described in detail below.
[0075] The intent generation module performs the following processing: It performs temporal modeling on the historical observation trajectories of each pedestrian in the same scene to obtain the temporal features of each pedestrian; it determines the local nearest neighbors of the target pedestrian at each observation time, constructs a relative interaction state based on the relative position and relative speed between the target pedestrian and each local nearest neighbor, calculates soft interaction weights, and uses the soft interaction weights to perform a weighted summation of the local nearest neighbor features to obtain neighborhood aggregation features; it generates social interaction correction features by combining the target pedestrian's own features and the difference between them and the neighborhood aggregation features, and cross-fuses the social interaction correction features with the temporal features of the target pedestrian to obtain social interaction features; it forms individual intent and social interaction intent from the temporal features of the target pedestrian and the social interaction features, and obtains the endpoint intent representation through gated residual fusion.
[0076] The intent generation module mainly consists of a temporal feature extraction module, a social interaction feature generation module, and a gated intent fusion module. The temporal feature extraction module performs temporal modeling on the historical observation trajectory of the target pedestrian, obtaining temporal features to characterize its own movement trend during the observation period, and providing a foundation for subsequent construction of individual intents. The social interaction feature generation module generates social interaction features based on the interaction relationships between the target pedestrian and surrounding pedestrians. At each historical observation moment, the temporal feature extraction module determines the target pedestrian's local neighbors based on the spatial distance between pedestrians, and constructs relative interaction states based on the relative positions and relative speeds of the target pedestrian and each local neighbor, updating the interaction relationships as the pedestrian's position changes. The relative interaction states are processed by a learnable weighted network to obtain soft interaction weights, which are used to weighted aggregate local neighbor features and form neighborhood aggregated features. The target pedestrian's own features and the differences between them and the neighborhood aggregated features are concatenated, and the concatenation result is input into a correction network to obtain social interaction correction features. These social interaction correction features are then cross-fused with the target pedestrian's temporal features to obtain the final social interaction features. Social interaction features characterize the impact of surrounding pedestrian interactions on the future destination of the target pedestrian and serve as the basis for constructing social interaction intentions. The gating intention fusion module generates individual intentions and social interaction intentions based on temporal features and social interaction features respectively. Gating weights are calculated from both, and the gating-regulated social interaction intention is superimposed as a residual correction term onto the individual intention to obtain the destination intention representation. This representation preserves the target pedestrian's own movement trend and incorporates social interaction information under gating regulation.
[0077] The pattern branch generation module is used to fuse the endpoint intent representation with multiple learnable pattern conditions to obtain multiple pattern features. Each pattern feature then predicts the corresponding candidate endpoint, forming a one-to-one "pattern feature - candidate endpoint" candidate branch. The pattern features are also passed to the trajectory decoding module as conditions for generating a complete future trajectory for the corresponding candidate branch. This ensures that the pattern-related features from the candidate endpoint prediction stage continue into the future trajectory generation stage, and constrains the same candidate branch in both stages.
[0078] The trajectory decoding module performs the following processing on each candidate branch: based on the historical conditions of the observed trajectory formation, and using the pattern features of the current candidate branch as the pattern conditions; weighting the historical conditions and the pattern conditions using fusion weights related to the future predicted location to construct a future motion prior feature sequence; combining the observed trajectory feature sequence, the future motion prior feature sequence, and the candidate endpoint features of the current candidate branch into an input sequence, and inputting the input sequence into the spatiotemporal representation network for decoding to generate a complete future trajectory.
[0079] The trajectory decoding module includes a future motion prior construction module and a future trajectory generation module. The future motion prior construction module constructs a future motion prior feature sequence for the current candidate branch. The temporal features and social interaction features of the observed trajectory are aggregated and fused to form historical conditions, used to represent the motion trend and social interaction state during the observation phase; the pattern features of the current candidate branch serve as pattern conditions, jointly constructing the future motion prior feature sequence with the historical conditions. The fusion weights related to the future prediction position adjust the relative strength of the two types of conditions along the future time step, so that prior features closer to the observation end emphasize historical motion trends, and prior features closer to the prediction end emphasize the pattern conditions of the current candidate branch, balancing short-term motion continuity and long-term pattern consistency. The future trajectory generation module combines the observed trajectory feature sequence, the future motion prior feature sequence of the current candidate branch, and the candidate endpoint features into a joint feature sequence, inputting it into the spatiotemporal representation network of the trajectory generation branch for modeling and decoding to obtain the corresponding complete future trajectory. In one implementation, using the candidate endpoint as a reference, the predicted relative bias is added to the candidate endpoint to obtain the position of the generated trajectory at the endpoint time, enhancing the consistency between the trajectory and the candidate endpoint. Each candidate branch executes this process to obtain... A candidate future trajectory.
[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting pedestrian trajectories, characterized in that, It includes the following steps: Temporal modeling was performed on the historical observation trajectories of each pedestrian in the same scene to obtain the corresponding temporal features; The local neighbors of the target pedestrian are determined at each observation time. The relative interaction state is constructed based on the relative position and velocity between the target pedestrian and each local neighbor. The soft interaction weight is calculated, and the local neighbor features are weighted and summed using the soft interaction weight to obtain the neighborhood aggregation feature. By combining the target pedestrian's own features and the difference between them and the neighborhood aggregation features, a social interaction correction feature is generated and cross-fused with the temporal features of the target pedestrian to obtain the social interaction feature; The individual intent and social interaction intent of the target pedestrian are formed by combining the temporal characteristics and social interaction characteristics to obtain the destination intent representation; The endpoint intent representation is fused with multiple learnable pattern conditions to obtain multiple pattern features, and the corresponding candidate endpoints are predicted to form multiple candidate branches. For each candidate branch, based on the historical conditions formed by the observed trajectory, the pattern features of the current candidate branch are used as pattern conditions, and the historical conditions and pattern conditions are weighted by the fusion weights related to the future predicted position to construct a sequence of prior features of future motion. The observed trajectory feature sequence, the future motion prior feature sequence, and the candidate endpoint features of the current candidate branch are combined into an input sequence, which is then input into a spatiotemporal representation network for decoding to generate a complete future trajectory.
2. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The method for obtaining the neighborhood aggregation feature specifically includes the following steps: At each observation time, a preset number of nearest neighbors are selected for the target pedestrian based on the Euclidean distance between their absolute coordinates, forming a local nearest neighbor set. The relative position vector and relative velocity vector between the target pedestrian and each local nearest neighbor are concatenated to obtain the relative interaction state vector; The relative interaction state vector is input into the weight network and softmax normalized within the local nearest neighbor set to obtain the soft interaction weight. The neighborhood aggregation feature is obtained by weighting and summing the features of each neighbor in the local nearest neighbor set using the soft interaction weight.
3. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The target pedestrian's own features are concatenated with the difference obtained by subtracting the neighborhood aggregation features from the target pedestrian's own features, and the concatenation result is input into the correction network to obtain the social interaction correction features; Using the temporal characteristics of the target pedestrian as the query and the social interaction correction feature sequence as the key and value, the social interaction features are obtained through cross-attention calculation.
4. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The method for obtaining the endpoint intent representation specifically includes the following steps: Attention is converged on the temporal features and social interaction features of the target pedestrian to obtain the individual intent vector and the social interaction intent vector. The individual intent vector and the social interaction intent vector are concatenated and then gating vectors are generated using the Sigmoid function. The social interaction intention vector is gated using the gating vector, and the gating result scaled by a preset weight coefficient is added to the individual intention vector to obtain the endpoint intention representation.
5. The pedestrian trajectory prediction method as described in claim 1, characterized in that, By concatenating the endpoint intention representation with each mode condition embedding vector and inputting them into a shared mode mapping network, multiple mode feature vectors are obtained; each mode feature vector is input into a shared coordinate prediction network to obtain the corresponding candidate endpoint coordinates.
6. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The method for constructing the prior feature sequence of future motion specifically includes the following steps: For the first time period excluding the end time within the prediction period For each future moment, obtain the learnable basic feature vector and temporal position embedding vector at that moment; Construct the motion prior feature vector for this future moment using the following formula: ; in, For the first Learnable basic features corresponding to each future moment For the first Embedding of the temporal position of a future moment, The historical conditions formed by the temporal characteristics and social interaction characteristics of the target pedestrian. This represents the pattern feature vector of the current candidate branch; and The fusion weights, which are related to the future predicted location, are used to adjust the influence of historical conditions and model conditions on the first... The strength of the influence of prior features of future motion.
7. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The calculation formula for the social interaction correction feature is as follows: ; In the formula, For the social interaction correction features, To calibrate the network; The characteristics of the target pedestrian; This is a neighborhood aggregation feature.
8. The pedestrian trajectory prediction method as described in claim 4, characterized in that, The formula for calculating the intended endpoint is as follows: ; In the formula, The intended endpoint is represented as follows. For the individual's intention, For preset weighting coefficients, Let be the gate vector. This indicates element-wise multiplication. The stated social interaction intention; The formula for calculating the candidate endpoint is: ; ; in, For the pattern features, For pattern mapping networks, For the first Conditional embedding of patterns, The candidate endpoint. For coordinate prediction networks, This is a candidate branch.
9. The pedestrian trajectory prediction method as described in claim 1, characterized in that, The network parameters of the prediction method are obtained through a phased training process, which includes: In the short-term prediction training phase, it uses real trajectories as input to train the spatiotemporal representation network to establish the ability to model time series. In the endpoint prediction training phase, the intention fusion and pattern branch generation parts are trained on the basis of the spatiotemporal representation network, and differential constraints that suppress pattern feature convergence and suppress excessive concentration of candidate endpoints are introduced. In the complete trajectory prediction training phase, the training endpoint prediction branch and trajectory generation branch are combined.
10. A pedestrian trajectory prediction system, characterized in that, The system is used to execute the pedestrian trajectory prediction method as described in any one of claims 1-9, the system comprising: The intent generation module performs the following processes: It performs temporal modeling on the historical observation trajectories of each pedestrian in the same scene to obtain the temporal features of each pedestrian; it determines the local nearest neighbors of the target pedestrian at each observation time, constructs a relative interaction state based on the relative position and relative speed between the target pedestrian and each local nearest neighbor, calculates soft interaction weights, and uses the soft interaction weights to perform a weighted summation of the local nearest neighbor features to obtain neighborhood aggregation features; it generates social interaction correction features by combining the target pedestrian's own features and the difference between them and the neighborhood aggregation features, and cross-fuses the social interaction correction features with the temporal features of the target pedestrian to obtain social interaction features; it forms individual intent and social interaction intent from the temporal features of the target pedestrian and the social interaction features, and obtains the endpoint intent representation through gated residual fusion. The pattern branch generation module is used to embed and fuse the endpoint intent representation with multiple learnable pattern conditions to obtain multiple pattern features, and predict the corresponding candidate endpoints by each pattern feature to form a one-to-one "pattern feature - candidate endpoint" candidate branch. The trajectory decoding module performs the following processing on each candidate branch: based on the historical conditions of the observed trajectory formation, and using the pattern features of the current candidate branch as pattern conditions; weighting the historical conditions and the pattern conditions using fusion weights related to the future predicted location to construct a future motion prior feature sequence; combining the observed trajectory feature sequence, the future motion prior feature sequence, and the candidate endpoint features of the current candidate branch into an input sequence, and inputting the input sequence into a spatiotemporal representation network for decoding to generate a complete future trajectory.
