一种改进的无人机对地小目标检测方法

An improved small object detection method on drone

无人机拥有空中视野良好,可监测范围广等优势,被广泛地应用于实际的目标检测任务中,由于无人机距离地面较远,任务中频繁出现小型目标检测效果不佳,虚检率和漏检率较高的情况。针对以上问题,提出一种改进的无人机对地小目标识别方法。本文基于YOLOV3卷积神经网络,首先建立一个无人机航拍数据集,并使用维度聚类方法设计合适的锚框,其次将广义交并比应用于网络的坐标损失函数中,替代原本的和方差损失,最后将YOLOV3网络4倍降采样特征图与经过上采样的8倍降采样特征图进行拼接,建立新的4倍降采样的目标检测层。实验结果表明,相比于YOLOV3,应用广义交并比的网络的平均精确度均值提高了3.4%,应用改进的YOLOV3网络平均精确率均值提高了8.2%,其中行人类小目标的平均精确率提高了10.2%,改进的检测方法对无人机平台下的小目标检测效果有所提升。

The drone has the advantages of good aerial vision and wide monitoring range.It is widely used in practical target detection tasks.Because the drone is far away from the ground,the small target detection is not effectively in the task,and the false detection rate and missing detection rate are high.To solve the problems above,An improved small object detection method on Drone is proposed.Based on the YOLOV3 convolutional neural network,this paper first establishes a Drone aerial photography dataset and uses the dimensional clustering method to design the appropriate anchor boxes.Secondly,Generalized Intersection over Union is applied to the coordinate loss function of the network to replace the original sum of squares dueto error method.Finally,the 4×downsampling feature map of the YOLOV3 network is spliced with the upsampled 8×downsampling feature map to establish a new 4×downsampling object detection layer.The experimental results show that compared with YOLOV3,the mean Average Precision of the network using the Generalized Intersection over Union is increased by 3.4%,and the mean Average Precision of the improved YOLOV3 network is increased by 6.2%,and the mean Average Precision of the small human target is improved by 10.2%,the improved detection method has improved the detection of small targets under the drone platform.