基于改进 YOLOv5-DeepSORT 算法的公路路面病害智能识别.

Aiming at the problems that target detection algorithms are prone to miss detection and misdetection of highway pavement lesions in multi-scale unmanned aerial vehicle (UAV) images and that the same lesion is repeatedly detected in consecutive frames, we propose a pavement detection method with the ability of lesion re-recognition. By introducing the true width-height loss and aspect ratio to improve the performance of the loss function, utilizing the CA (Coordinate attention) mechanism to improve the recognition ability of the model in complex backgrounds, introducing the initial features of the model into the feature fusion network to improve the robustness of the model in detecting multi-scale pavement lesions, and constructing a second-level detection mechanism based on the DeepSORT (Multi-target Tracking Algorithm for Target Detection) to realize the re-recognition of lesions. The experimental results show that the average detection accuracy of the model mAP reaches 89. 19%, which is 3. 11% higher than that of the benchmark model; the F1 score is 0. 851 4, which is 2. 49% higher than that of the original model; at the same time, it also outperforms the mainstream target detection algorithms; and the counting accuracy of the lesions in the drone image reaches 91. 38%, which is 25. 86% higher than that of the pre-improvement model, which provides accurate and real-time lesion data for the road pavement inspection and maintenance. [ABSTRACT FROM AUTHOR]

针对目标检测算法在多尺度无人机 (UAV) 图像中对公路路面病害容易出现漏检误检及同一病害在连续帧图片 中被重复检测的问题,提出一种具有病害重识别能力的路面检测方法。通过引入真实宽高损失与纵横比以提升损失函数性 能,利用 CA(Coordinate attention)注意力机制提升模型在复杂背景下的识别能力,将模型初始特征引入特征融合网络提升模 型检测多尺度病害的鲁棒性,构建基于 DeepSORT(目标检测的多目标跟踪算法) 的二级检测机制实现对病害的重识别与统 计。试验结果表明,模型平均检测精度mAP 达到89. 19%,较基准模型提升了3. 11%;F1 分数为0. 851 4,较原模型提升了 2. 49%;同时也优于主流目标检测算法;在无人机影像下病害计数精度达到91. 38%,较改进前提升25. 86%,为公路路面检测 与养护提供精确实时的病害数据。 [ABSTRACT FROM AUTHOR]