An Improved YOLOv5s Algorithm and Its Application to Real-Time Target Detection for Autonomous Driving

Fast detection speed is an advantage of the YOLOv5s algorithm as well as simple deployment, easy scalability, and lightweight design. Among its many applications are video surveillance, automatic driving, industrial inspection, and so on. However, it still has the problems of insufficient detection accuracy, abnormal model fitting, and bias accumulation. Therefore, this study offers an enhanced YOLOv5s real-time target identification method, SCSYOLOv5s, addressing the problems of insufficient detection accuracy, aberrant model fitting, etc. Aiming at the problem of ambiguity of the object described by the aspect ratio of the CIoU loss function, the SIoU loss function is introduced, which enhances model training speed and inference accuracy. Aiming at the problems of memory leakage, insufficient computational accuracy, and bias accumulation in the SPPF pyramid architecture, to enhance the model’s ability to perceive details, the CAM attention mechanism is implemented. Aiming at the difficulties of slow convergence speed of AdamW optimizer and overfitting in the pre-training phase resulting in aberrant model fitting conditions, SGD optimizer is added to further increase the detection accuracy and training efficiency of the model. The proposed model is applied to the VOC2007 dataset for validation, and the results show that the SCS-YOLOv5s model proposed in this paper has a high detection accuracy in the validation set P of 0.8638, R of 0.8417, and $\mathbf{m A P} \text{@} 0.5$ of 0.9043. The model has a high detection accuracy, and the validity of the improved strategy is verified. Compared with other models, the identification precision of the model in this research is enhanced by $0.58 \% \sim 5.55 \%$, and the superiority of the model is verified. The SCS-YOLOv5s model suggested in this article has numerous advantages, including excellent precision in detection, quick convergence speed, and favorable detail perception, which is applicable to the field of real-time detection and recognition of automatic driving.