YOLO-O: An Improved YOLO-Based Framework for Vehicle Detection

For road safety, human factors and ergonomics play a vital role, influencing the object detection. To automate the object detection, several algorithms like R-CNN, SSD, R-FCN, and YOLO have been used for this task. While these models are effective for many object detection applications, detecting tiny objects on roads remains challenging due to low precision. Developing a lightweight and robust model capable of high-precision detection of tiny objects is essential. This study proposes a YOLO-O model, built on an enhanced YOLOv7 model, to detect tiny vehicles. A residual network with three convolutional layers and Squeeze and Excitation Network (SENet) replaces the Efficient Layer Aggregation Network (ELAN) in YOLOv7. Coordinate attention modules are added to improve recognition accuracy, and the SIoU regression loss enhances bounding box prediction. Moreover, an 80 × 80 detection head improves the recognition of small vehicles. The model, tested on COCO, achieved superior performance with an mAP of 94.9%, outperforming YOLOV2, YOLOV3, and YOLOV5. Moreover, the proposed model performed significantly on cross-validation using the PASCAL Visual Object Classes and LISA datasets. This system demonstrates the method's practicality and robustness for tiny vehicle recognition, offering a foundation for developing automated object detection and tracking.