Title: Lightweight object detection algorithm for automatic driving scenarios

Authors: Zhou Tu; Ming Chen

Addresses: College of Computer and Information Technology, China Three Gorges University, Hubei, Yichang, China; Hubei Key Laboratory of Intelligent Vision-Based Monitoring for Hydroelectric Engineering, Hubei, Yichang, China ' College of Computer and Information Technology, China Three Gorges University, Hubei, Yichang, China; Hubei Key Laboratory of Intelligent Vision-Based Monitoring for Hydroelectric Engineering, Hubei, Yichang, China

Abstract: Aiming to solve the problems of large model size, unbalanced detection speed and detection accuracy of traditional target-detection algorithms in autonomous driving scenarios, a lightweight YOLO algorithm is proposed based on YOLOv4. First, the lightweight network Mobilenetv1 was used to replace the original YOLOv4 feature-extraction network and a Depth-Wise Cross-Stage Part module (DW-CSP) was proposed to improve the detection speed. Then, a new Lish activation function was designed and the K-means++ clustering algorithm was used to regenerate prior frames of different scales. Finally, the FocalLoss loss function was introduced instead of the cross-entropy loss function. Experiments show that compared with the YOLOv4 algorithm, the improved YOLO algorithm improves the detection accuracy by 1.72% and the detection speed by 53%, and the model size is reduced by about four times. The algorithm is more in line with the target-detection requirements of autonomous driving scenarios.

Keywords: automatic driving; object detection; lightweight network; YOLOv4; activation function.

DOI: 10.1504/IJWMC.2024.136584

International Journal of Wireless and Mobile Computing, 2024 Vol.26 No.1, pp.74 - 82

Received: 23 Dec 2022
Accepted: 27 Jun 2023
Published online: 07 Feb 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article