BiGA-YOLO: A Lightweight Object Detection Network Based on YOLOv5 for Autonomous Driving
Jun Liu, Qiqin Cai, Fumin Zou, Yintian Zhu, Lyuchao Liao, Feng Guo
Abstract
Object detection in autonomous driving scenarios has become a popular task in recent years. Due to the high-speed movement of vehicles and the complex changes in the surrounding environment, objects of different scales need to be detected, which places high demands on the performance of the network model. Additionally, different driving devices have varying performance capabilities, and a lightweight model is needed to ensure the stable operation of devices with limited computing power. To address these challenges, we propose a lightweight network called BiGA-YOLO based on YOLOv5. We design the Ghost-Hardswish Conv module to simplify the convolution operations and incorporate spatial coordinate information into feature maps using Coordinate Attention. We also replace the PANet structure with the BiFPN structure to enhance the expression ability of features through different weights during the process of fusing multi-scale feature maps. Finally, we conducted extensive experiments on the KITTI dataset, and our BiGA-YOLO achieved a [email protected] of 92.2% and a [email protected]:0.95 of 68.3%. Compared to the baseline model YOLOv5, our proposed model achieved improvements of 1.9% and 4.7% in [email protected] and [email protected]:0.95, respectively, while reducing the model size by 15.7% and the computational cost by 16%. The detection speed was also increased by 6.3 FPS. Through analysis and discussion of the experimental results, we demonstrate that our proposed model is superior, achieving a balance between detection accuracy, model size, and detection speed.