Real-time Traffic Congestion Detection with SIGHTA Regression Network
Long Jiang, Yatao Wang, Ying Zhao · 2019
This work proposes a novel SIGHTA regression network to detect traffic congestion in real-time via video stream. The network inherits the architecture of YOLOv3 [1] which uses Darknet53 backbone. Referring to SPP network [2], the network structure has been adjusted so that more objects can be detected. Additionally, a novel regression algorithm and loss function are presented to regress quadrilateral objects. Experimental results demonstrate that the SIGHTA network significantly outperforms the existing object detection methods by mAP. Meanwhile the speed is faster.