A YOLOv3-Based Learning Strategy for Vehicle-Thrown-Waste Identification.

At present, throwing objects from car windows is increasingly becoming a major illegal conduct that pollutes the urban environment and affects the city's image. Apart that, the rubbish casually thrown from car windows is left in the motorway, which seriously threatens the lives of sanitation workers. Moreover, it is time-consuming and laborious to manage through manual monitoring. In response to this phenomenon, we propose an improved yolov3-based real-time target detection for throwing objects from car windows. In order to solve the problem of deep hierarchy in the YOLOv3 network and improve the accuracy and speed of small target detection, the following improvements are made: Firstly, random feature sampling and interpolation are added to the residual blocks of the YOLOv3 backbone network to reduce the computational effort of the CNN network while maintaining high performance. Secondly, the Soft NMS is changed to Matrix NMS to speed up the identification of the best candidate boxes. Finally, a residual layer was removed to improve the accuracy of small target detection. The experimental study evaluates and demonstrates a number of improvements in accuracy and speed compared to current mainstream target detection methods.