改进YOLOX的轻量化航拍目标检测算法

A lightweight algorithm for aerial image object detection based on YOLOX

摘要:

针对小型无人机在巡逻航拍中的应用，提出了一种改进的轻量化目标检测算法，有效解决巡逻过程中空地无线传输信道和机载端计算能力双重受限的难题；该算法在YOLOX算法的基础上，首先利用Mobilenetv2代替CSPDarknet骨干网络作为特征提取网络，降低了模型参数量和计算量，提高目标检测实时性；其次为了弥补轻量化带来的检测精度下降，考虑检测目标框的长宽比引入CIOU定位损失函数，提升目标定位的精度；同时为了平衡训练过程中的正负难易样本，引入Focal Loss置信度损失函数提升模型的检测性能；基于VisDrone2019-DET数据集实验表明，改进后算法模型参数量降低了56.2%，计算量降低了52.5%，在检测精度没有明显下降情况下单张图片推理时间减少了41.4%；最后，将改进后的算法部署到Nvidia Jetson Xavier NX机载端，测得模型检测帧率可达22FPS，改进后算法满足巡逻任务的应用需求。

Abstract:

A lightweight object detection algorithm is proposed for small unmanned aerial vehicle (UAV) patrol applications, which can effectively solve the dual constraints of wireless transmission channel and on-board computing resource. Firstly,based on YOLOX algorithm, Mobilenetv2 network is used as feature extraction network to reduce the number of model parameters and improve the speed of object detection. Secondly, CIOU loss function is used instead of IOU function to improve object positioning accuracy. Thirdly, Focal Loss function was introduced to balance the positive and negative difficult samples in training to improve the performance of the model. Experiments based on VisDrone2019-DET dataset show that the improved algorithm reduces the number of model parameters by 56.2%, the calculation amount by 52.5%, and the inference time of a single image by 41.4% without significant decrease in detection accuracy. Finally, the improved algorithm is deployed to the Nvidia Jetson Xavier NX, and the model detection frame rate can reach 22FPS, which meets the application requirements of patrol tasks.