Traffic Monitoring

Roadways in the city of Virginia Beach have been ranked consistently as top crash locations in the US. Existing infastructure such as vision-based camera, thermal cameras, radars, Dedicated Short Range Communication Road Side Units and other have been used primarily to control traffic flow. This study focuses on leveraging existing infrastructure, specifically vision based cameras to provide insights and identify dangerous traffic scenarios using computer vision (adapted from Safe-D).

My involvement in this project was to build an automated computer vision system to identify cars, locate them and track them through videos collected by the City of Virginia Beach cameras. I used popular object detectors such as Faster RCNN, Cascade RCNN, RetinaNet, RefineNet and others to locate the cars in each frame. Those videos suffer however from low image quality and high noise. For this reason, I performed training and specifically fine-tuning to improve the performance. After detecting the objects, I applied an object tracker, called Deep SORT, to track each object in consecutive frames.

Demo of our work is below. Detections are with red boxes. Tracked objects are with different color and with numbers. The tracker is able to filter out the noise from the low-confidence detections and improve results significantly even in the difficult conditions of low image quality and noise.