Search

Aiming at real-time detection of UAVs, small UAV targets are easily missed and difficult to detect in complex backgrounds. To maintain high detection performance while reducing memory and computational costs, this paper proposes the SEB-YOLOv8s detection method. Firstly, the YOLOv8 network structure is reconstructed using SPD-Conv to reduce the computational burden and accelerate the processing speed while retaining more shallow features of small targets. Secondly, we design the AttC2f module and replace the C2f module in the backbone of YOLOv8s with it, enhancing the model's ability to obtain accurate information and enriching the extracted relevant information. Finally, Bi-Level Routing Attention is introduced to optimize the Neck part of the network, reducing the model's attention to interfering information and filtering it out. The experimental results show that the mAP50 of the proposed method reaches 90.5 % and the accuracy reaches 95.9 % , which are improvements of 2.2 % and 1.9 % , respectively, compared with the original model. The mAP50-95 is improved by 2.7 % , and the model's occupied memory size only increases by 2.5 MB, effectively achieving high-accuracy real-time detection with low memory consumption. [ABSTRACT FROM AUTHOR]

The vulnerable road users (VRUs), being small and exhibiting random movements, increase the difficulty of object detection of the autonomous emergency braking system for vulnerable road users AEBS-VRUs, with their behaviors highly random. To overcome existing problems of AEBS-VRU object detection, an enhanced YOLOv5 algorithm is proposed. While the Complete Intersection over Union-Loss (CIoU-Loss) and Distance Intersection over Union-Non-Maximum Suppression (DIoU-NMS) are fused to improve the model's convergent speed, the algorithm also incorporates a minor object detection layer to increase the performance of VRU detection. A dataset for complex AEBS-VRUS scenarios is established based on existing datasets such as Caltech, nuScenes, and Penn-Fudan, and the model is trained using migration learning based on the PyTorch framework. A number of comparative experiments using models such as YOLOv6, YOLOv7, YOLOv8 and YOLOx are carried out. The results of the comparative evaluation show that the proposed improved YOLO5 algorithm has the best overall performance in terms of efficiency, accuracy and timeliness of target detection. [ABSTRACT FROM AUTHOR]