Road Target Detection Based on Lightweight Improved YOLOv5l

• Published in: 2023 8th International Conference on Information Systems Engineering (ICISE) pp. 244 - 248
• Main Authors: Xu, Haoran, Yu, Weiwei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 23.06.2023
• Subjects: artificial intelligence; Computational modeling; Feature extraction; lightweight network; machine Learning; Object detection; Real-time systems; road targets; Road traffic; Roads; Solid modeling; target detection; YOLOv5l
• ISSN: 2643-7309
• DOI: 10.1109/ICISE60366.2023.00058

In the current stage of autonomous driving technology, road target detection is a key problem for performing related operations. However, the most of the target detection models are not suitable or difficult to be applied in practical real-time road target detection due to the problems of the large parameter computation and the long discriminative time. To address this series of issues, In this paper, we propose a method based on the lightweight improvement of YOLOv5l used into the target detection of road target. Firstly, the Slim-Neck structure is introduced into the original model's Neck to decrease the volume of operations and increase the accuracy of the original model; secondly, the framework which is used in the backbone network of YOLOv5l is altered by the PP-LCNet network structure to reduce the large volume of operations required by the model in the feature extraction of images and substantially increase the operating speed of the model; Finally, inspired by RFBNet, the SPPF layer of the original model is replaced by the RFB structure in RFBNet to prevent the loss of model accuracy in the case of a large reduction of floating point operations and to increase the original model's detection accuracy. According to the experimental results, the improved YOLOv5l model is compared with the original YOLOv5l model, the size of model is reduced to 43.33% of the original size, the operation parameters of model are reduced to 43.95% of the original size, the floating point operations of model are reduced to 28.23% of the original size, and the accuracy of model is improved by 2.46% to 95.77%. The result of experiments shows that the model which is proposed in this paper based on the lightweight improved YOLOv5l can be more efficiently and accurately applied in the practical scenarios for road target detection.