RE-YOLOv5: Enhancing Occluded Road Object Detection via Visual Receptive Field Improvements

Abstract

Road object detection technology is a key technology to achieve intelligent assisted driving. The complexity and variability of real-world road environments make the detection of densely occluded objects more challenging in autonomous driving scenarios. This study proposes an occluded object detection algorithm, RE-YOLOv5, based on receptive field enhancement to assist with the difficult identification of occluded objects in complex road environments. To efficiently extract irregular features, such as object deformation and truncation in occluded scenes, deformable convolution is employed to enhance the feature extraction network. Additionally, a receptive field enhancement module is designed using atrous convolution to capture multi-scale contextual information and better understand the relationship between occluded objects and their surrounding environment. Considering that the ordinary non-maximum suppression method in dense occlusion scenarios will incorrectly suppress the prediction box of the occluded object, EIOU was used to optimize the non-maximum suppression method. Experiments were conducted on two benchmark datasets, KITTI and CityPersons. The proposed method achieves a mean average precision (mAP) of 82.04% on KITTI, representing an improvement of 2.34% over the baseline model. For heavily occluded objects on CityPersons, the Log Average Miss Rate (MR−2) is reduced to 40.31%, which is a decrease of 9.65% compared to the baseline. These results demonstrate that the proposed method significantly outperforms other comparative algorithms in detecting occluded objects across both datasets.

Keywords

• object detection
• occlusion object detection
• receptive field
• YOLOv5