A study on attention-based LSTM for abnormal behavior recognition with variable pooling

Abstract

Behavior recognition is a well-known computer vision mobile technology. It has been used in many applications such as video surveillance, motion detection on devices, human-computer interaction and sports video, etc. However, most of the existing works ignored the depth and spatio-temporal information so that they resulted in over-fitting and inferior performance. Consequently, a novel framework for behavior recognition is proposed in this paper. In this framework, we propose a target depth estimation algorithm to calculate the 3D spatial position information of the target, and take this information as the input of the behavior recognition model. Simultaneously, in order to obtain more Spatio-temporal information and better handle long-term video, combining with the idea of attention mechanism, we propose a skeleton behavior recognition model which is based on spatio-temporal convolution and attention-based LSTM (ST-CNN & ATT-LSTM). The deep spatial information is merged into each segment, and the model focuses on the key information extraction, which is essential for improving behavior recognition performance. Meanwhile, we use a feature compression method based on variable pooling to solve the problem of inconsistent input sizes caused by multi-person behavior recognition, so that the network can flexibly recognize multi-person skeleton sequences. Finally, the proposed framework is evaluated with real-world surveillance video data, and the results indicate that our framework is superior to existing methods.