Vehicle Detection Using You Only Look Once V8 Based On Architecture Modification Method

Object detection remains challenge by significant constraints in computer vision, including difficulty in recognizing small, occluded, or contextually overlapping objects. Existing architectures often exhibit limitations in accuracy and speed, particularly in dynamic environments. This Research propose a modified YOLOv8 architecture method by adding the RepNCSPELAN4 Module layer to the head to detect to optimize detect layer on small and large layer vehicles on traffic scenes. The training process was compared to YOLOv8, and the Proposed Model. This study found that the adjustment of YOLOv8 architecture in the detection segment can affect the preparation results and provide much better detection results than the default YOLOv8, which only achieves the highest mAP value of 94 and poor detection of small objects. Specifically, it improves mAP0.5 from 0% to 0.30%, and mAP0.5:0.95 from 0% to 0.021% in small object detection. In real traffic scene situations, RepNCSPELAN4-YOLOv8 illustrates significant changes in various conditions, demonstrating its quality and effectiveness in identifying small objects even in challenging situations.