Enhanced Object Detection of Abnormal Light Based on Multi-scale Retinex with Chromacity Preservation

In autonomous driving scenarios, the lighting conditions can often change rapidly and dramatically, leading to overly bright or dark environments. These abnormal light conditions can negatively impact the accuracy of object detection, potentially causing accidents and posing a threat to pedestrian and driver safety. To address this issue, this paper presents a method to mitigate the effect of abnormal light intensity on object detection in autonomous driving scenarios. We apply Multi-scale Retinex with Chromacity Preservation (MSRCP) processing to abnormal light images based on the YOLO model in order to recover lost information. Our method was evaluated on the publicly available BDD 100k dataset, and the results showed a significant improvement in performance under abnormal lighting conditions, Compared to images without any processing, Yolo-v3 model improve 9.59 \(\%\) total mAP@[.5, .95], and Yolo-v4 model improves 10.86 \(\%\) .