An Improved Road Target Detection Approach Based on YOLO-v8

To address the challenges of high equipment demands, low detection accuracy, and high miss rates for overlapping and small targets in traditional road target detection models, we propose an improved detection method named YOLOv8-GPT. The C2f module in the backbone network is restructured with the inclusion of a lightweight attention mechanism, Triplet Attention, improving the model's feature extraction capabilities. Additionally, a new lightweight convolution accumulation technique, GSConv, is integrated into the Neck, reducing model burden while preserving accuracy and enhancing computational cost efficiency. Then, to better detect small vehicles in crowded traffic, a dedicated small target detection head has been added, aiming to capture detailed features and contextual cues effectively. Ultimately, the Wise-IoU, capable of adaptively adjusting weight coefficients, is employed as the loss function for the refined model, boosting the regression performance of bounding boxes and overall detection robustness. Experiments on the KITTI dataset show that the YOLOv8-GPT model significantly excels beyond the original, with a 3.5 percentage point increase in <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m$</tex> AP, a reduction of 1.3 million parameters, and superior performance in both volume and accuracy compared to other classic detection models, highlighting its strong practical applicability.