Interpretable Roadside 2D Object Detection Based on Axial Attention

Environmental perception plays a crucial role in the operation of autonomous vehicles. Recent studies have relied on single vehicle cameras for perception due to the lack of roadside view dataset, but this approach provides limited perception information. To address this limitation, there is a growing interest in studying roadside perception to complement vehicle-side perception information. However, existing methods face challenges in accurately recognizing and locating small objects or objects under occluding conditions in the roadside. And autonomous vehicles require millisecond-level decision-making for safe driving, yet many advanced object detection methods struggle to meet the need for high-speed reasoning. Additionally, deep learning models often make black-box decisions without providing explanations or understanding of their decision-making process. To tackle these issues, this study proposes an interpretable 2D object detection method for roadside based on the YOLO V8 with Axial Attention replacing the CNN network within the network architecture. The trained model is then input into HiResCAM to visualize the decision-making process of the roadside detection model. Experimental results on DAIR-V2X-I and Rope3D demonstrate that our proposed method not only improves mean Average Precision (mAP) but also enhances image processing speed by 1.5 to 2 times compared to existing methods. By visualizing the decision-making process of the roadside detection model, it make autonomous vehicle perception more transparent and credible.