Vehicle Target Detection Research Based on Enhanced YOLOv8

In light of the traditional target detection algorithms with inadequate accuracy and weak robustness in modern scenarios, e.g. autonomous driving, an improved YOLOv8-based vehicle target detection algorithm is proposed. Firstly, the attention mechanism (GAM) is implemented to optimize the backbone and neck networks based on YOLOv8, improving the network's feature extraction ability. Secondly, the EIoU loss function of the Focal high-quality anchor frames is added to replace the IoU loss function, accelerating network convergence. Lastly, the Hardswish activation function replaces the original network's Sigmoid function in the global network, enhancing target detection accuracy. The experiment findings reveal that the refined algorithm enhances precision by 11.8%, recall by 1.9%, and mean average accuracy by 4.8% in comparison to the YOLOv8 algorithm on the BDD100k dataset. Alongside this, the detection speed satisfies the real-time prerequisites of autopilot, indicating the efficiency of the upgraded algorithm on the in-house dataset.