Video Anomaly Detection using Inflated 3D Convolution Network
Dipali Koshti, Supriya Kamoji, Nehal Kalnad, Suyash Sreekumar, Shreya Bhujbal
Abstract
The infrastructure of every city is getting smarter day by day. This infrastructure provides us with vital information. There is a demand for a real-time system that can help identify crimes as soon as they happen, which is now possible with the rise in popularity of AI. The data captured through the surveillance system can contain anomalous and normal videos. We propose to build an anomalous event detection system by using weakly labelled training videos, and after the detection of such an activity, an alarm would be raised. A Deep residual learning framework called I3D-Resnet- 50 is used for feature extraction. This network is pre-trained on the Kinetics video action dataset. Our dataset consists of 13 distinct anomalies. Anomalous events are Robbery, Assault, Shooting, Burglary, Stealing, Arrest, Fighting, Shoplifting, Arson, Explosion, Vandalism, Abuse, Road Accident. The introduced method for video anomaly detection attains significant improvement in the results both in terms of accuracy and recall.