Suspicious Activity Detection for Security Cameras

Abstract

Human activity recognition, in recent years, has become a very important topic in artificial intelligence and computer vision. The need for automated surveillance in security cameras has grown since the human eye cannot detect suspicious activities accurately. Detection of suspicious activities and automatic reporting can prevent crime before it occurs, thus avoiding collateral damage. In this paper, we have classified human activities into two: abnormal and normal. Normal activities include sitting, walking, hand waving, etc. Abnormal activities include kicking, punching, pointing a gun, wielding a knife, etc. We achieve this classification by using convolutional and recurrent neural networks. First, the convolutional neural network is used to extract high level features from images. The final classification of the convolutional network is not considered, instead, the result of the last pooling layer is extracted and passed to the recurrent neural network to make the final prediction.