Abnormal Activity Detection from CCTV Camera Feeds

Abstract

Video surveillance is widely utilized in both public and private settings for observation and monitoring purposes, making it a prominent application of computer vision technology. It resulted in less human work put towards oversight. Homes, workplaces, hospitals, malls, parking lots, etc. places can use smart video surveillance systems. Due to its many uses, including the identification of criminal behavior, traffic accidents, and unlawful activities, abnormal detection in video surveillance is a well-liked study topic in computer vision. Abnormal activity means any behavior or event that is not considered normal in a particular situation. For example, in CCTV footage, abnormal activity could include unusual or suspicious actions like theft, vandalism, or violence. The definition of abnormal activity can change based on the context and the objectives of the surveillance system. In this paper, we have done a literature survey on abnormal activity detection and experimented with detecting violence in videos using a convolutional neural network on the Real-Life Violence Situations 1 (RLVS) dataset. Here we have implemented three different model architectures that incorporate spatial and temporal features for violence detection. The first model combines MobileNetV2 with an RNN layer, while the second model employs an LSTM layer instead. The third model based on residual LSTM. The fourth model represents a fusion of LSTM and RNN, using the strengths of both.