基于无人机图像的城市道路停车巡检方法

An Inspection Method of Urban Road Parking Based on UAV Image

快速准确的对城市道路停放车辆进行巡检对于城市智慧管理具有重要意义。针对当前巡检方法的低效率、高成本和不准确,研究了1种基于无人机图像的巡检方法。首先,为满足全天候巡检,采用光照度增强算法对暗光条件的图像进行增强,同时采用去模糊算法对模糊图像质量进行改善。针对YOLOv5现有算法检测精度不高,实时性不强等问题,改进网络为使模型收敛速度更快,采用Focal-EIOU Loss优化损失函数;为了可以更好的适应不同大小和形状的目标,采用C2F模块替代C3模块,使用不同大小的卷积核提取特征;为了提高网络的鲁棒性及抗干扰能力,通过添加SimAM注意力机制,不增加网络模块参数且能预测特征图的3D注意力权值;采用CARAFE算子进行上采样增加感受野,全面利用特征图的语义信息。实验结果表明:改进后YOLOv5模型的准确率提高了5.1%,召回率提高了5.9%,平均精度mAP值提高了3.6%。其次,采用字符识别网络SVTR对车牌号进行识别,仅通过单个视觉模型就能完成特征提取和文本转录2个任务。通过无人机场工程应用平台进行试验,试验结果表明:该巡检方法可以准确、快速且智能完成巡检,准确度达到90%,检测速度达到170帧/s,基本满足巡检精度和实时性要求。

Efficient and accurate inspection of parked vehicles on urban roads is of significant importance for smart city management.Addressing the issues of low efficiency,high cost,and inaccuracy in current inspection methods,a drone-based image inspection approach is investigated.To enable all-weather inspection,an illumination enhance-ment algorithm is employed to boost images captured under low-light conditions,while a deblurring algorithm is uti-lized to improve the quality of blurred images.To tackle the limitations of the existing YOLOv5 algorithm,includ-ing insufficient detection accuracy and real-time performance,several modifications are introduced.The Fo-cal-EIOU Loss function is optimized to accelerate model convergence.The C3 module is replaced with the C2F module,utilizing varying sizes of convolutional kernels to extract features,enhancing adaptability to targets of dif-ferent sizes and shapes.Furthermore,the SimAM attention mechanism is incorporated to improve the network's ro-bustness and anti-interference capability,predicting 3D attention weights for feature maps without increasing model parameters.The CARAFE operator is adopted for upsampling to expand the receptive field,comprehensively lever-aging semantic information from feature maps.Experimental results demonstrate that the modified YOLOv5 model achieves 5.1%increase in accuracy,5.9%improvement in recall,and 3.6%enhancement in mean Average Precision(mAP).Secondly,the SVTR character recognition network is utilized to identify license plate numbers,accomplish-ing both feature extraction and text transcription tasks within a single vision model.Finally,field tests conducted on a drone-based engineering application platform reveal that this inspection method can accurately,rapidly,and intelli-gently complete inspections,achieving an accuracy of 90%and a detection speed of 170 frames per second,essen-tially meeting inspection precision and real-time requirements.