Video Anomaly Detection Using Pre-Trained Deep Convolutional Neural Nets and Context Mining

Monitoring of CCTV is readily accessible today, but software is unable to accurately detect aberrant behaviour, such as shoplifting or sudden running, which is referred to as an anomaly. The existing techniques trained the model to recognize abnormal behaviours by using all normal images, but they were strongly dependent on a huge number of training pictures. The author of this work uses three distinct pretrained CNN models/algorithms, dubbed anomaly detection, object detection, and tracking, to get around the problems. The pre-trained models exhibit high efficiency in accurately extracting features, which are subsequently retrained using the De-noising Auto Encoder model (DAE). After DAE has trained on normal images, it will alert users with red bounding boxes if it detects any aberrant frames in the movie. The UCSD pedestrian dataset, which is used in this proposed work to train the model, shows people walking, which is considered normal behaviour. If, however, a person arrives on a bicycle or a tree falls, the model detects this as aberrant behaviour, also known as anomaly detection. The distinction between normal and pathological behaviour, known as spectral and temporal behaviour—spectral denoting place, and temporal denoting TIME—is extracted through the application of context mining. In order to carry out this experiment, we used the UCSD images dataset, which is taken from movies.