Multi-Modality Abnormal Crowd Detection with Self-Attention and Knowledge Distillation

Deep Neural Networks (DNNs) have become a promising solution for detecting abnormal human behaviors. However, building an efficient DNN model in terms of both computational cost and classification accuracy is still a challenging problem. Furthermore, there are limited existing datasets for abnormal behavior detection, and each focuses on a certain context. Therefore, a DNN model trained on a certain dataset will be adaptive for a particular context and not suitable for others. This study proposes a DNN framework with efficient attention and Knowledge Distillation (KD) mechanisms. Attention units capture key information from multiple RGB, optical flow, and heatmap inputs. KD is applied to scale down model size. Experiments were performed on several benchmark datasets, examining both AUC and accuracy. The results show that the proposed framework outperformed other state-of-the-art methods in detection accuracy. Furthermore, the trade-off between detection performance and computational cost was also addressed by the proposed framework with KD.