Video anomaly localization using modified faster RCNN with soft NMS algorithm

Localization of anomalies in surveillance videos is a critical component of smart and intelligent surveillance systems. The goal of anomaly detection is to automatically detect the presence of anomalies in a short amount of time. The proposed system developed an efficient and improved faster RCNN-based system for accurate detection of anomalies. The modified version of Faster RCNN extracts the features at different levels by using a feature pyramid network and is able to detect small-scale anomalies by adding a Soft NMS algorithm. The proposed model is experimentally evaluated using three benchmarked datasets UCSD Ped1, UCSD Ped2 and Avenue, and gets better detection performance. Finally, a comparison study with different YOLO series is conducted, and our system outperforms them.