A Comparative Study of Faster R-CNN Models for Anomaly Detection in 2019 AI City Challenge

Linu Shine, Anitha Edison, Jiji C. V.; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 306-314

Abstract


Traffic anomaly detection forms an integral part of intelligent traffic monitoring and management system. Timely detection of anomalies is crucial in providing necessary assistance to accident victims. Track 3 of 2019 AI city challenge addresses traffic anomaly detection problem. We propose an unsupervised method to tackle this problem. Proposed system consists of three stages. The first stage is a background extraction stage which isolates the stalled vehicles from moving vehicles. An anomaly detection is the second stage that identifies the stalled vehicles in the background and finally anomaly confirmation module confirms anomaly and determines the start time. We have used faster RCNN (FRCNN) with Inception v2 and ResNet 101 to detect stalled vehicles and confirm possible anomalies. A comparative study shows that FRCNN with Inception v2 gives superior performance.

Related Material


[pdf]
[bibtex]
@InProceedings{Shine_2019_CVPR_Workshops,
author = {Shine, Linu and Edison, Anitha and , Jiji C. V.},
title = {A Comparative Study of Faster R-CNN Models for Anomaly Detection in 2019 AI City Challenge},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}