ADAM Optimizer Based Convolutional Auto Encoder for Detecting Anomalies in Surveillance Videos

In recent years video surveillance systems have gained much attention by installing them in public places such as railway stations, airports, traffic signals, banks, streets, institutions, etc. Main goal of the surveillance system is to accurately detect video anomalies in real time thereby reducing anomalous activities. Anomalous activities are abnormal patterns that occur in the video that are not similar to regular sequences. Various anomalies like, fight, fire accident, riot, gun fire, handling dangerous weapons, etc. Though many methods have been developed to detect the anomalies, many fail to detect the anomalies due to various conditions like environmental changes, occlusions, complex human behaviors etc. To overwhelmed these issues and to improve the accuracy of video anomaly detection a novel method is proposed using Convolutional Auto Encoder. The proposed method reconstructs the given input data and generates the loss for the given video sequences; further the Reconstruction loss is given to a score estimator to estimate the normality of sequence. To increase the detection rated and reduce the computational time ADAM optimizer is utilized for optimizing the auto encoder network. Further a comparative analysis is made on state-of-the-art techniques, with proposed method which shows that a significant improvement in accuracy is achieved.