Research on Lightweight Obstacle Detection Model Based on Improved YOLOv5s
Yuting Ma, Guodong Xu · 2024
The paper intends to investigate the accuracy and lightweight by improving the YOLOv5s model to enhance the sensing ability of campus AGVs on eight types of dynamic obstacles, and effectively solve the problem that it is difficult to balance the real-time and accuracy of its target detection. On the basis of YOLOv5s model, GhostConv is introduced to replace the ordinary convolution, which reduces the number of parameters and computation of the model without affecting the detection accuracy, and at the same time, it adopts the weighted bidirectional feature pyramid BiFPN and the edge regression loss function EIoU to obtain the feature maps with richer feature information, which improves the model generalization ability and detection performance. The results show that the improved model significantly reduces the number of parameters and computation of the model with little effect on the mean average precision mAP, which decreases by 16% and 13%, respectively, compared with the YOLOv5s model, while the inference speed is also improved by 28%, realizing the balance between lightweight and detection accuracy, and the obtained lightweight network model is useful for the embedded devices with weak computational capability is more friendly.