Research on a Lightweight Panoramic Perception Algorithm for Electric Autonomous Mini-Buses

Abstract

Autonomous mini-buses are low-cost passenger vehicles that travel along designated routes in industrial parks. In order to achieve this goal, it is necessary to implement functionalities such as lane-keeping and obstacle avoidance. To address the challenge of deploying deep learning algorithms to detect environmental information on low-performance computing units, which leads to difficulties in model deployment and the inability to meet real-time requirements, a lightweight algorithm called YOLOP-E based on the YOLOP algorithm is proposed. (The letter ‘E’ stands for EfficientNetV2, and YOLOP-E represents the optimization of the entire algorithm by replacing the backbone of the original model with EfficientNetV2.) The algorithm has been optimized and improved in terms of the following three aspects: Firstly, the YOLOP backbone network is reconstructed using the lightweight backbone network EfficientNet-V2, and depth-wise separable convolutions are used instead of regular convolutions. Secondly, a hybrid attention mechanism called CABM is employed to enhance the model’s feature-representation capability. Finally, the Focal EIoU and Smoothed Cross-Entropy loss functions are utilized to improve detection accuracy. YOLOP-E is the final result after the aforementioned optimizations are completed. Experimental results demonstrate that on the BDD100K dataset, the optimized algorithm achieves a 3.5% increase in mAP50 and a 4.1% increase in mIoU. During real-world vehicle testing, the detection rate reaches 41.6 FPS, achieving the visual perception requirements of the autonomous shuttle bus while maintaining a lightweight design and improving detection accuracy.

Keywords

• deep learning
• model lightweighting
• attention mechanism
• depth-wise separable convolution
• YOLOP
• electric autonomous mini-bus