Road Traffic Object Detection System for Severe Weather Scenarios Based on YOLOX

Abstract

Road traffic object detection is the core perceptual task in the current field of intelligent transportation. The You Look Only Once (YOLO) series algorithms are widely used for real-time detection due to their end-to-end inference, fast speed, high accuracy, and ease of deployment. However, existing public datasets are mostly collected in sunny weather, and harsh conditions such as rainy, foggy, snowy, and nighttime can lead to image quality degradation, blurred object edges, reduced contrast, and insufficient lighting, resulting in a significant decrease in detection performance. In response to the above issue, this paper uses YOLOX as the basic model, introduces SE module and small object detection head method to improve the detection ability of the model in complex environments, and verifies the effectiveness of the proposed method on the Adverse Conditions Dataset with Correspondences (ACDC) dataset. The above improvements resulted in a mAP value of only 21.86% on the ACDC dataset, a decrease of about 0.41% compared to the baseline method, but an increase of 1.8% in small object AP values and 2% in car class detection. In the future, further optimization can be attempted from the perspective of changing the introduction method.