Aimed at overcoming issues like limited object types, missed detection, and false positives in existing models, an improved object detection algorithm for autonomous driving based on YOLOv8s is proposed. Ordinary convolutions in the YOLOv8s backbone are replaced with RepConv (Re-parameterization Convolution) to enhance target perception while reducing computational load and memory consumption, thereby improving model efficiency. Additionally, an efficient multi-scale attention (EMA) mechanism is introduced after the neck’s C2f block to strengthen feature attention and accelerate model convergence. A P2 detection head is also added to improve small object detection capabilities. Finally, the WIoU (Wise-IoU) loss function, featuring a dynamic non-monotonic focusing mechanism and gradient gain allocation strategy, is employed to boost overall detector performance. On a manually labeled Car dataset, the improved model achieved mAP50 and mAP(50-95) scores of 81.2% and 58.4%, respectively, 1.5% and 1.2% higher than the original YOLOv8s model. Precision and recall are improved by 1.9% and 0.8%, and the parameter count is decreased from 11.14M to 10.87M. The proposed modules increase detection accuracy while reducing parameter count, making the model more suitable for autonomous driving applications.