This study addresses the issues in smart traffic management and autonomous driving by presenting an improved YOLOv8s-based model for real-time road object recognition. The model is optimized by utilizing the COCO dataset and methods including He initialization, layer-wise adaptive rate scaling (LARS), stochastic weight averaging (SWA), and adaptive gradient clipping (AGC). Furthermore, a cosine annealing learning rate schedule, which uses warm restarts and non-maximum suppression (NMS), increases accuracy and reduces false positives. The identification of tiny objects is further refined by post-processing innovations, such as the inclusion of squeeze-and-excitation (SE) blocks and feature pyramid networks (FPN). The YOLOv8s model outperforms 12 other cutting-edge models, according to experimental data, with 97.9% accuracy, 96.8% recall, and a 97.3% F1 score. Additionally, the model generates a low root mean square error (RMSE) of 0.188, greatly surpassing the results of earlier iterations such as YOLOv7 and Faster R-CNN. These findings demonstrate the model's dependability in precisely identifying objects on the road, which makes it ideal for real-time applications like autonomous cars and traffic management. By striking a balance between accuracy and computing economy, the suggested technique provides a convincing way to improve road object recognition.