GC-YOLOX: Privacy Small Object Detection Algorithm

Abstract This study addresses the challenges in autonomous driving scenarios, including the unclear definition of privacy targets, the propensity for missed detections on small road objects, as well as the generally low detection efficiency. To tackle these issues, we have constructed a privacy target detection dataset and proposed a YOLOX-based algorithm for autonomous driving perception. Firstly, the privacy target detection dataset was developed in accordance with the 'Several Provisions on Automotive Data Security Management.' Secondly, an Efficient Multi-scale Residual Attention (EMA) mechanism was constructed and integrated with the concept of residual connections to enhance the network's capacity for processing feature information. Thirdly, we introduced the Global Contextual Information Fusion (GCIF) structure to enrich the feature information within images, thereby improving the detection of small targets and refining classification and regression tasks. Fourthly, the Enhanced Intersection over Union (EIOU) Loss function was employed to further refine the algorithm's regression capabilities, which in turn, elevated the model's overall performance. Finally, in the experimental evaluation, our proposed algorithm was compared with other mainstream algorithms using both the newly constructed privacy target detection dataset and the established KITTI dataset. The comparative analysis demonstrated that our algorithm outperformed others in terms of detection performance. Specifically, on the privacy target detection dataset, the mean Average Precision (mAP), mAP at 0.5 Intersection over Union (mAP@0.5), and mAP at a small object scale (mAP@s) were 47.2%, 86.4%, and 41.7%, respectively. On the KITTI dataset, these metrics achieved 64.9%, 92.6%, and 55.4%, respectively. These results indicate a significant enhancement in the detection performance of the YOLOX algorithm for privacy targets and small road objects.