Research on the Detection Algorithm for Abnormal Crowd Behaviors Based on an Enhanced SlowFast Model
Yueping Peng, Hexiang Hao, Tongtong Zhou, Baixuan Han, Wenji Yin · 2024
Aiming at the issues of inaccurate detection of behavioral classification in crowd abnormal behaviors detection algorithms and large model computation capacity, a crowd abnormal behaviors detection algorithm based on the improved SlowFast model is proposed. By changing the convolutional structure of the network to a Depthwise-separable convolution, adding a hybrid domain attention mechanism, and inserting Dropout to inhibit the overfitting phenomenon, we can enhance the network's precision of detection and classification of behaviors while ensuring the computational capacity. Experiments show that the method in this paper can detect and identify five types of behaviors in public places: normal, scattered, speeding up in the same direction, sudden gathering, and group fighting. Compared with the original SlowFast model and other algorithms, the precision is improved, and the number of detected frames per second can be up to 40.5 fps, which meets the real-time requirements, providing a new idea for the determination and classification of abnormal behaviors of crowds.