| Abstract |
Cleaning vehicles, vital in municipal projects, automated via unmanned driving, boost urban cleaning efficiency and sustainability, cutting costs and minimizing environmental impact. Environmental perception is key to their automation, differing from passenger cars. Cleaning vehicles often encounter issues such as low cleaning efficiency, high energy consumption, and suboptimal cleaning results when operating in complex and variable scenarios. The automated cleaning vehicle uses an environmental perception system to identify waste and provides vital information for controlling its cleaning devices, boosting both cleaning and energy efficiency. However, existing object detection algorithms struggle with challenges such as background interference and the small size of road debris, leading to limited detection capabilities. This paper proposes an environmental perception solution specifically tailored for the operational context of cleaning vehicles. The solution uses a binocular camera system, integrating stereo vision technology with object detection algorithms to identify and locate road debris. Attention mechanisms are incorporated to enhance model detection accuracy. An additional layer for small object detection is implemented to improve the ability to detect smaller targets, increasing the mean average precision by 5.1% and the recall rate by 6.61%. These improvements could lead to more accurate and efficient automated cleaning systems in urban environments, with far-reaching implications for the future of smart cities.To validate the feasibility of the proposed solution, tests are conducted using a road debris dataset and real-vehicle application experiments. The experimental results demonstrate that the system effectively performs debris identification and localization tasks on cleaning vehicles. © 2025 Elsevier Ltd |
