Abnormal Behavior Recognition Based On Deep Learning For Video

The identification of abnormal behavior under video surveillance is a topic of practical significance,which is of great significance to the formation of intelligence in the security industry.With the development of deep learning in computer vision,the use of deep learning to solve abnormal behavior recognition under video surveillance has become one of the research hotspots in recent years.A comprehensive network for video anomaly behavior recognition based on deep learning is composed of a detection network and an identification network.The detection network consisting of the C3 D feature extraction network and the multi-instance learning fully connected network for feature scoring uses a semi-supervised learning method,first tagging at the video level instead of the frame level,and then using the normal and abnormal video together to identify the network learning abnormal behavior,and finally detect when an abnormal behavior occurs.In the identification network,the 2D-DenseNet is merged with the characteristics of the C3 D network,and the DenseC3 DNetwork is formed by the migration learning to perform the behavior recognition of the video stream.Since the normal behavior is high in the video,the proportion of abnormal behavior is low.The behavior recognition of normal behavior not only causes waste of resources,but also reduces the recognition accuracy of the recognition network.The detection network can detect the occurrence of abnormal behavior in the video and mark the occurrence time period,discard the frame of normal behavior and integrate the frame of abnormal behavior into the behavior recognition network for abnormal behavior recognition,so tha the identification network does not need to identify normal behavior as a type of behaviors.The experimental results show that the detection network has no missing detection examples for the non-intensive,unoccluded,and well-lit abnormal behavior video,but there are instances in which the normal behavior is falsely reported as abnormal behavior.The identification network has 79%recognition accuracy on the data set of UCF101-Crime,and there are some examples of unrecognized or identification errors.