The human action recognition technology has developed rapidly in recent years. The technologies of RNN and 3D convolution based on posture information and video frame information respectively have achieved high accuracy using various data sets, however, both of them have shortcomings in the field of abnormal behavior recognition. The definition of abnormal behavior needs to consider not only the action type simply, but also the environmental information comprehensively, so there are limitations in using RNN only based on posture information. Due to the input characteristics, action recognition technology based on 3D convolution is more related to environmental information and group behavior information, it cannot locate the action time accurately. This paper proposed an abnormal behavior recognition framework based on P3D and LSTM. The framework used pre-trained P3D to extract environmental features, and adopted pre-trained LSTM to extract individual action features to help system for time positioning, finally apply ranking model to classify abnormal behaviors after combining environmental features with action features. When training LSTM model, a regression network was added to enhance its time positioning ability. The experiment showed that the proposed framework based on P3D and LSTM has a greater improvement in the recognition accuracy and time positioning than only using 3D convolution technology or LSTM technology, and can accurately recognize abnormal behaviors.
原文链接
分享
引用
