With the rapid development of intelligent transportation systems, accurately perceiving and understanding pedestrian behavior has become the key to ensuring road safety. However, the diversity of pedestrian occlusion and behavior in complex traffic...
With the rapid development of intelligent transportation systems, accurately perceiving and understanding pedestrian behavior has become the key to ensuring road safety. However, the diversity of pedestrian occlusion and behavior in complex traffic scenarios poses significant challenges to environmental perception systems. This paper focuses on this issue and conducts systematic research on pedestrian detection and behavior recognition based on human posture skeleton sequences, aiming to develop an algorithm solution that combines high precision and high real-time performance. In the field of 2D human pose estimation, this paper proposes the lightweight network IDP-Net. This network takes the high-resolution network as the baseline and innovatively designs the IDP basic module and the peer feature fusion module. While ensuring the accuracy of pose estimation, it significantly reduces the number of parameters and computational complexity of the model, meeting the strict requirements of the intelligent driving system for the real-time performance of the algorithm. To address the challenge of pose estimation in severely occlusive scenarios, this paper proposes a 3D human pose estimation CTP model based on multiple perspectives. This model operates directly in the 3D voxel space. By projecting 2D joint features from multiple perspectives onto a unified three-dimensional space, it effectively solves the problem of joint occlusion in a single perspective. The CTP model adopts a strategy of first detecting the center point of pedestrians and 3D bounding boxes, and then regressing the center points of each joint, avoiding the complex process of matching joint points and performing well in multi-person occlusion scenarios. At the level of behavior recognition, this paper proposes a multi-stream adaptive attention subgraph convolution method based on graph convolutional neural networks. By traversing the sequence of depth-first trees to construct subgraphs, adaptively learning the association weights among various body parts, and embedding the ECA channel attention module, the recognition accuracy of pedestrian behavior has been significantly improved. The multi-stream framework is adopted to integrate the physical structure information and motion information of the human body, further enhancing the algorithm's ability to discriminate complex behaviors. To support algorithm research, this paper also constructs a pedestrian behavior dataset for intelligent transportation scenarios, including a singleview dataset based on on-board cameras and a multi-view dataset based on multi-view visual sensors at intersections, providing an important data foundation for training and validating behavior recognition algorithms. The experimental results show that the series of methods proposed in this paper have achieved competitive performance on public benchmark datasets, effectively enhancing the accuracy and robustness of pedestrian perception in complex traffic scenarios, and providing reliable technical support for the practical application of intelligent transportation systems.