An efficient and lightweight small target detection framework for vision-based autonomous road cleaning

The development of machine vision technology provides a feasible way for intelligent vehicles used for road cleaning to achieve automation, improve cleaning efficiency, and reduce energy consumption. This paper proposes an efficient and lightweight small target detection framework, in which a cascade model is formed by cascading a refined model for category correction to the back-end of a general model for garbage detection to perform the high-precision detection of garbage in road scenes, and a road segmentation model is proposed to determine the operating area of the cleaning vehicle. For the road segmentation model, a category-based loss function is proposed to improve the recall rate of difficult categories such as sidewalk and foreground. Besides, we reparameterize the network structure in the inference phase of the model, thus achieving a lightweight model to improve computational efficiency and reduce the need for the use of computational resources. The experimental results show that the proposed road segmentation model achieves a good trade-off between accuracy and speed compared to the state-of-the-art model. For the cascade model, it can realize the category correction of some targeted garbage with low confidence and optimize the detection effect of road garbage. The experimental results demonstrate that compared with using YOLOv5s alone, the average recall rate of our proposed cascade model for road garbage is improved from 77.8% to 81.4%, the average accuracy rate is improved from 69.5% to 83.1%, and the detection performance for small-sized garbage is also better.