Research on passengers behavior recognition method in public transport vehicles based on efficient 3D CNN

To solve the problem of behavior recognition algorithm performance degradation caused by space limitation, difficulty in obtaining passenger data, and behavior complexity in public transport vehicles, this paper constructs the P-bus dataset for passenger abnormal behavior. Additionally, we proposed the Star 3D CNN (S3D) algorithm, based on a high-dimensional, nonlinear feature space mapping element-wise multiplication module (Starblock-3D) and a local self-similarity attention mechanism module (Simneck-3D). It addresses the characteristics of skeletal keypoints modalities by extracting intricate and nonlinear spatial information from high-dimensional skeletal keypoints. It also generates weights through local self-similarity to apply weighting across both spatial and channel dimensions, thereby effectively capturing motion characteristics among skeletal keypoints. The results reveal that S3D enhances the accuracy by 38.26% and reduces the number of parameters by 95.4% compared to the most recent algorithms for recognizing passenger behaviors. Without pre-training parameters, S3D achieves improvements of 1.91% in top1_acc on UCF101 and 2.55% on HMDB51, along with gains of 1.78% and 2.55% in mean_class_acc, respectively, compared to the Slowonly algorithm. The S3D network demonstrates a high degree of effectiveness and accuracy in detecting abnormal passenger behaviors within public transportation systems.