Improved YOLOv8 complex road scenes object detection algorithm

Abstract:

To enhance small object detection in complex road senses, an improved YOLOv8 algorithm is introduced. Firstly, the quadruple down-sampling branch is added to improve the learning ability of the network for small object features. Secondly, a SPPF-BRSA module based on Bi-Level Routing Spatial Attention (BRSA) is designed to remove irrel-evant regions in a query adaptive manner, which effectively reduces the interference of complex background on detection performance. In addition, the C2fDynamic module is used in the neck of YOLOv8 to strengthen the feature expression ability of the model by dynamically selecting the convolution kernel. Finally, the Wise-IoU v3 loss function is used to obtain more accurate detection results by adjusting the gradient gain distribution. The experimental results show that on Huawei SODA10M dataset, the improved algorithm improves precision, recall, F1 score and mAP50 by 4.6%, 4.1%, 4.3% and 5.4% respectively compared with the original algorithm.

Key words:YOLOv8; deep learning; road scenes; attention mechanism; loss function