Road sludge detection and identification based on improved Yolov3

In order to deal with the challenge of the identification task of road sludge under real scenes, we propose a novel detection for road sludge detection. It combines the road sludge features extracted by the residual network with the feature maps of various scales. The swish activation function is used in the network, and GIoU-loss is used as the loss of position regression. The Improved-YOLOv3 experiment on our self-built road sludge datasets shows superior performance in speed and accuracy and it also realized real-time detection and recognition. Specially, the detection time of each frame is 45ms on the GPU (GTX 1080ti) acceleration. Furthermore, it achieved a 96.7% precise rate and a 63.5% recall rate.