Traffic Speed Estimation from Surveillance Video Data: For the 2nd NVIDIA AI City Challenge Track 1

Abstract

Estimating traffic flow condition is a tough but beneficial task. In Intelligent Transportation System (ITS), many applications have been done to collect and analyze traffic data. However, the surveillance video data are still only used for engineer's manual check. To better utilize this data source, traffic flow estimation from surveillance camera should be explored. This study uses Faster Regional Convolutional Neural Network (Faster R-CNN) with ResNet 101 as the backbone to achieve multi-object detection. Then a tracking algorithm based on histogram comparison is applied to link objects across frames. Finally, this study uses warping method to convert vehicle speeds from the pixel domain to the real world. The results show that estimating vehicle speed at intersection is more challenging than in uninterrupted flow.