Abstract: To address the issue that the existing methods for recognizing abnormal behaviors of elevator passengers in complex scenarios have insufficient understanding of behavioral characteristics, resulting in low recognition accuracy, an improved PPTSM model (CMTSM) recognition method is proposed. Firstly, to address the problem that the ResNet50 network fails to capture behavioral characteristics adequately, a multi-scale dilated attention mechanism is introduced, enabling the model to integrate multi-scale semantic information and effectively capture the feature information of abnormal behaviors. Secondly, to enhance the feature information of human posture behaviors, a vertical depth point convolution fusion module is proposed. This module integrates high-level semantic features and low-level detail features, allowing the network to extract more comprehensive features. Finally, experiments are conducted on a custom dataset of four abnormal behaviors (door-pulling, jumping, kicking, and falling) and public datasets (UCF24, UCF101, HMDB51, and Something-Something-v1). The experimental results show that compared with the original PPTSM model, the proposed model has significant improvements in evaluation metrics such as accuracy, recall rate, precision, and F1-score, and can meet the actual requirements for detecting abnormal behaviors of elevator passengers in complex scenarios.