Anomaly Behavior Detection in Crowd via Lightweight 3D Convolution

Detecting abnormal behaviors in large crowds is a highly challenging problem. In practical scenarios, there are two problems. One of the problems is the imbalance in the number of training samples collected through monitoring always results in overfitting, which is mainly reflected in the lack of samples of specific abnormal behavior for training. The other problem is that edge computing devices in practical scenarios cannot load models with many parameters, while most deep learning methods need excessive parameters to obtain good performance. In this article, a lightweight model is proposed to detect anomalous behaviors, which combines memory singular value decomposition and attention mechanism in 3D convolution. Truncated singular value decomposition is adopted to prevent overfitting, in which a cross-entropy loss is introduced into the loss function. The integrated 3D separable convolution decomposes 3D convolutions into 2D and 1D channels, which can effectively reduce the number of parameters. Compared to traditional deep learning models, experimental results have shown the proposed model is superior to others in performance on several benchmark datasets.