Search

Accurately predicting the future trajectory of pedestrians is critical for tasks such as autonomous driving and robot navigation. Previous methods for pedestrian trajectory prediction dealt with social interaction and pedestrian movement factors either concurrently or sequentially, neglecting the link between them. Therefore, a spatial–temporal interaction-aware graph convolution network (STIGCN) is proposed for pedestrian trajectory prediction. STIGCN considers the correlation between social interaction and pedestrian movement factors to achieve more accurate interaction modeling. Specifically, we first constructed spatial and temporal graphs to model social interactions and movement factors. Then, we designed the spatial–temporal interaction-aware learning to utilize the spatial interaction features of each moment to assist the temporal interaction modeling and utilize the temporal interaction features of each pedestrian to assist the spatial interaction modeling, resulting in more accurate interaction modeling. Finally, a time-extrapolator pyramid convolution neural network (TEP-CNN) is designed to jointly estimate the two-dimensional Gaussian distribution parameters of future trajectories by combining the prediction features from multiple layers. Experimental results on two benchmark pedestrian trajectory prediction datasets show that our proposed method outperforms existing methods in terms of average displacement error and final displacement error and achieves more accurate predictions for pedestrian motions such as convergence and encounter. [ABSTRACT FROM AUTHOR]

• Pedestrian trajectories were jointly modeled using both social interactions and movement factors. • Asymmetric 3D convolution was used to further process the adjacency matrices of the spatial and temporal graphs, so as to realize the fusion of spatial-temporal information, so that the model could learn the continuity of social interaction and the movement factors. • The RSigmoid function was designed to assign weights to the adjacency matrices. While holding the integrity of the interaction information, appropriate weights were given to the micro-interactions to achieve a reasonable setting of interaction weights. • The U-TCN module achieved better trajectory prediction effects by combining the information of the front and back temporal convolution network (TCN) layers. The pedestrian trajectory prediction remains challenging due to its uncertainty and interference from surrounding pedestrians. There are two deficiencies in previous pedestrian trajectory prediction methods: 1. The temporal correlation of social interaction and the movement factors of groups are ignored; 2. Unreasonable interaction weight allocation. In order to eliminate these two deficiencies, a reasonably dense graph convolution network (RDGCN) was developed in this study. Spatial and temporal graphs were first constructed to model social interactions and movement factors. Then, asymmetric three-dimensional (3D) convolution was employed for the fusion of spatial-temporal information to capture the temporal correlation of social interactions and the movement factors of groups. The RSigmoid function was designed to assign interaction weights and to make the setting of interaction weight more reasonable. Finally, a U-TCN module was designed to estimate two-dimensional Gaussian distribution parameters of the future trajectories. On the ETH and UCY datasets, the proposed method outperformed versus other models in terms of average displacement error and final displacement error, and it was capable of predicting complex social behaviors and movement factors. [ABSTRACT FROM AUTHOR]

Accurately understanding and predicting the future trajectories of pedestrians around autonomous vehicles is a major challenge. Due to the uncertainty and diversity of pedestrian trajectories, current research focuses on the multimodality of trajectory prediction. Currently, multimodal pedestrian prediction methods mainly use deep generative networks to generate mutually independent trajectories through independent sampling strategy, thus focusing on modes with a large number of samples. Moreover, deep generative networks are black-box models that are entirely data-driven and cannot effectively explain trajectory generation. In addition, current research lacks an effective measure for each trajectory. Therefore, in our paper, we propose a physics constrained pedestrian trajectory prediction with probability quantification. We divide multimodal trajectory prediction into two phases: trajectory generation phase and trajectory selection phase. During the trajectory generation phase, rather than directly predicting trajectories, we incorporate a differential constraint module to ensure that the generated trajectories adhere to pedestrian motion laws. This approach enhances the interpretability of the model prediction process. Subsequently, we generate a set of candidate trajectory proposals with specific correlations through the correlated sampling module. During the trajectory selection phase, a probability selection module is proposed to establish explicit probability distributions for the candidate trajectory proposals and to select the proposal with the highest probability as the final output. Extensive experiments on a public real-world pedestrian trajectory dataset show that our proposed model exhibits significant advantages over existing models in multimodal trajectory prediction, which not only effectively reduces the error but also mitigates mode collapse. Moreover, we provide probabilities for each trajectory to better capture the model uncertainty and provide more comprehensive information for downstream decision-making systems. • Implemented differential constraints to regulate pedestrian behavior. • Proposed a correlation sampling method to mitigate modal collapse. • We propose an efficient probabilistic measurement mechanism. [ABSTRACT FROM AUTHOR]

Autonomous vehicles offer significant advantages for transportation systems, particularly in enhancing traffic safety. To achieve this goal, it is crucial to comprehensively understand and predict the future trajectories of pedestrians in proximity to autonomous vehicles. Many contemporary approaches for predicting pedestrian trajectories heavily rely on neural networks, especially recurrent neural networks. However, these approaches do not explicitly incorporate the dynamics of pedestrian movement and instead rely on data-driven black-box models. Consequently, these models may fall short in terms of interpretability and fail to adhere to the fundamental principles of kinematics. In response to these limitations, our work introduces an innovative model for pedestrian trajectory prediction grounded in neural differential constraints. We aim to investigate temporal changes in pedestrian state variables, such as position and speed, using neural networks. During the prediction process, the output of the neural network is governed by differential equations. This approach ensures that the generated trajectories align with the fundamental principles of physics, harnessing the combined power of neural networks and physics-based pedestrian motion models. Furthermore, our research endeavors to develop a cohesive framework that seamlessly integrates pedestrian movement patterns with the influence of ego-vehicles, while also considering potential destinations to inform future trajectory planning. We conducted extensive experiments on two publicly available real-world datasets to assess the effectiveness of our model in enhancing prediction accuracy and providing coherent explanations of pedestrian motion, comparing it to state-of-the-art methods. • Differential equations can restrict trajectories to adhere to physical principles. • Pedestrian-vehicle movement is converted to image coordinate system for modeling. • Enhance prediction accuracy by precisely estimating a pedestrian's potential goal. [ABSTRACT FROM AUTHOR]

Pedestrian trajectory prediction in first-person view is an important support for achieving fully automated driving in cities. However, existing pedestrian trajectory prediction methods still have significant shortcomings in terms of pedestrian trajectory diversity, dynamic scene constraints, and dependence on long-term trajectory prediction. We proposes a non-probability sampling network based on pedestrian trajectory anomaly recognition (ADsampler) to predict multiple possible future pedestrian trajectories. First, by incorporating pose and optical flow information, ADsampler models the multi-dimensional motion characteristics of pedestrians based on observed trajectory information and discriminates trajectory states. The sampling range in the Gaussian latent space is determined based on the recognition results. Next, velocity and yaw information of the car are introduced to model the car's motion state. A subtraction fusion network is employed to remove redundant image feature constraints in highly dynamic scenes. Finally, ADsampler utilizes a novel trajectory decoding network that combines the position encoding capability of GRU with the long-term dependency capturing ability of Transformer to decode and predict the fused features. we evaluate our model on crowded videos in the public datasets JAAD, PIE, ETH and UCY. Experiments demonstrate that the proposed method outperforms state-of-the-art approaches in prediction accuracy. • We propose a pedestrian trajectory anomaly recognition network. • we develop a subtraction fusion network. • We devise a trajectory prediction module GRU-Transformer. [ABSTRACT FROM AUTHOR]