Enhanced Transfer Learning-Based CNN for Abnormal Human Activity Detection in Video Surveillance Using Spatial-Temporal Features

Video surveillance (VS) is essential in today’s environment. Artificial intelligence (AI) including ML, and DL brought too much technological advancement to the surveillance system. The combination of these technologies assists in distinguishing distinct suspicious acts and behaviors from the real-time surveillance video. Human behavior is unpredictable; therefore, it is challenging to determine whether a behavior is suspicious or not. Thereby, we develop a new model for recognizing the abnormal actions of humans in VS. Initially, the input frames are preprocessed via Improved Wiener Filtering (IWF). As the next step, segmentation is done using the Improved SegNet model (ISegNet). Further, spatial and temporal (S&T) features, Improved Motion Estimation (ME), Color Feature, motion boundary SIFT (MoBSIFT) and Local Gradient Increasing Pattern (LGIP) features are extracted. Finally, the detection of abnormal actions of humans takes place via Improved Transfer Learning based CNN (ITL-CNN). The outcomes from ITL-CNN include Abuse, arrest, arson, assault, road accidents, robbery and shooting.