Key points are not available for this paper at this time.
The historical mission of transportation personnel is to build a strong transportation nation. The most important component of an intelligent transportation system is vehicle detection and analysis, which involves gathering information and monitoring the status of vehicles using advanced technologies to achieve effective allocation of transportation resources. In recent years, methods based on convolutional neural networks have made rapid advancements in the field of image processing. Addressing the challenges in small object detection, this paper proposes an improved algorithm for small object detection based on YOLOv3. Building upon the single-stage object detection algorithm YOLOv3, several improvements are made. Firstly, a modified K-means clustering algorithm is used to analyze and generate anchor boxes, providing an improved evaluation metric. Secondly, addressing the weak feature extraction capability and low detection accuracy of the backbone network for small objects, a lightweight network called EfficientNet is employed. EfficientNet-B6 is selected as the backbone network for experimental analysis based on the characteristics of the dataset. The proposed algorithm is validated using public datasets and compared with the improved Faster R-CNN algorithm, demonstrating that the improved YOLOv3 is suitable for real-time object detection. The detection accuracy achieved is 86.8%, with a detection speed of 39.5 frames per second on a 16GB GPU.
Wang et al. (Tue,) studied this question.