Jiang Shuhai, Luo Bowen, Jiang Haoyue, Zhou Zhongkai, Sun Shangjie
School of Mechanical and Electronic Engineering, Nanjing Forestry University, Nanjing, 210037, Jiangsu, China.
Institute of Intelligent Control and Robotics (IICR), Nanjing Forestry University, Nanjing, 210037, Jiangsu, China.
Sci Rep. 2024 Jan 11;14(1):1127. doi: 10.1038/s41598-024-51868-0.
The urban street is a congested environment that contains a large number of occluded and size-differentiated objects. Aiming at the problems of the loss of the target to be detected and low detection accuracy resulting from this situation, a newly improved algorithm, based on YOLOv4, DCYOLO is proposed. Firstly, a Difference sensitive network (DSN) is introduced to extract the edge features of objects from the original image. Then, assign the edge features back to increase the edge intensity of the object in the original image and ultimately improve the detection performance. Secondly, the feature fusion module (CFFB) based on context information is introduced to realize the cross-scale fusion of shallow fine-grained features and deep-level features, to strengthen the cross-scale semantic information fusion of feature maps and eventually improve the performance of object detection. At last, in the network prediction part, the SIOU loss function replaces the original CIOU loss function to improve the convergence speed and accuracy of object detection. The experiments based on MS COCO 2017 and self-made datasets show that, compared with the YOLOv4, the detection accuracy of DCYOLO models is greatly improved with an increase of 9.1 percentage points in AP and 10.4 percentage points in AP. Compared with YOLOv5x and Faster R-CNN, DCYOLO shows higher accuracy and better detection performance. The experiment result proves that the DCYOLO algorithm can adapt to the dense object detection requirements in the congested environment of urban streets.
城市街道是一个拥挤的环境，包含大量被遮挡且大小各异的物体。针对这种情况导致的待检测目标丢失和检测精度低的问题，提出了一种基于YOLOv4的新改进算法DCYOLO。首先，引入差异敏感网络（DSN）从原始图像中提取物体的边缘特征。然后，将边缘特征反馈回去以增强原始图像中物体的边缘强度，最终提高检测性能。其次，引入基于上下文信息的特征融合模块（CFFB），实现浅层细粒度特征和深层特征的跨尺度融合，加强特征图的跨尺度语义信息融合，最终提高目标检测性能。最后，在网络预测部分，用SIOU损失函数取代原来的CIOU损失函数，以提高目标检测的收敛速度和精度。基于MS COCO 2017和自制数据集的实验表明，与YOLOv4相比，DCYOLO模型的检测精度有了很大提高，AP提高了9.1个百分点，AP提高了10.4个百分点。与YOLOv5x和Faster R-CNN相比，DCYOLO显示出更高的精度和更好的检测性能。实验结果证明，DCYOLO算法能够适应城市街道拥挤环境下的密集目标检测需求。
