Introduction: As the key technology for autonomous driving, object detection methods have been widely reported in research publications and patents. The paper aims to address current challenges in traffic object detection, specifically insufficient precision for vehicles and poor performance in occluded object detection within complex traffic environments. To this end, it proposes YOLOv11mAEA, an enhanced algorithm based on the modified YOLOv11m. Method: Firstly, an Attention-based Intra-scale Feature Interaction (AIFI) module is introduced to achieve high-quality cross-scale feature interaction. Secondly, an Efficient Multi-Scale Plus Convolution (EMSPConv) module is incorporated to enhance multi-scale object detection capabilities. Finally, an Auxiliary Head (Aux Head) module is integrated into the network to provide additional gradient signals during training, thereby achieving overall network optimisation. The algorithm is trained and evaluated on the public Cityscapes dataset. Results: The evaluation framework employs Precision (P), Recall (R), mean Average Precision (mAP), and parameter count to comprehensively assess algorithm performance. When the AIFI module is introduced alone, the precision and mAP of the algorithm are improved by 1.2% and 1.6% respectively. When the EMSPConv module is introduced alone, the mAP of the algorithm is improved by 0.6%. When the Aux Head module is introduced alone, the precision and mAP of the algorithm improve by 0.3% and 1.0%, respectively. Ablation experiments results demonstrate that the YOLOv11mAEA algorithm achieves an accuracy of 76.1%, recall of 50.4%, and mAP of 58.7%, representing respective improvements of 3.7% in precision, 0.3% in recall, and 2.4% in mAP compared to the baseline algorithm. Furthermore, the detection performance for occluded objects has also been enhanced. Discussion: This study validates the performance of its proposed algorithm, which is based on YOLOv11m with three enhancements, under identical settings. Results may vary across parameters. Future work will focus on extensive validation on other large-scale datasets. Conclusion: Comparative experiments with mainstream algorithms reveal that the proposed method exhibits significant advantages in both precision and mAP metrics, demonstrating an enhanced capability to meet the detection requirements of complex traffic environments.
You et al. (Tue,) studied this question.