基于单目的轻量化目标识别系统设计

Lightweight Design of Target Recognition System Based on Monocular Vision

摘要:

为了满足微型无人飞行器实时控制的要求,文章对Yolo-V4单阶段图像检测算法进行了轻量化设计,并基于单目摄像头设计了实时目标识别系统。该轻量化算法首先通过优化输入网络的分辨率和卷积层参数等方法,将主干网络由73层压缩到16层；然后通过简化残差结构和跨级结构,解决深层网络梯度消失问题,减少梯度信息重复；最后自下而上将网络浅层和深层特征进行了精简融合。采用K均值聚类算法训练轻量化模型以优化预选框参数,提高算法对特定目标的检测精度。测试表明,改进后的轻量化算法对特定目标的预测精度为98%,召回率为92%,均值平均精度达90.70%,单张图片检测时间为7.8ms,对视频的处理速度可达125.6f/s,满足微型无人飞行器的实时应用要求。

Abstract:

To meet the requirement of real-time control of micro-unmanned aerial vehicles, a light-weighted Yolo-V4 single-stage image detection algorithm was designed, and a real-time target recognition system was designed based on a monocular camera. The light-weighted algorithm compresses the backbone network from 73 layers to 16 layers by improving the resolution of the input network and the parameters of the convolutional layer. And then by simplifying the residual and cross-level structures, the problem of deep network gradient disappearance was revolved, and the gradient information duplication was reducing. Finally, the shallow and deep features of the network were streamlined and merged from bottom to top. The K-means clustering algorithm was used to train the light-weighted model to improve the parameters of the pre-selection box and improve the detection accuracy of the algorithm for specific targets. Test results show that the proposed algorithm can reach a prediction accuracy of 98% for a specific target, a recall rate of 92%, a mean average accuracy of 90.70%, a single image detection time of 7.8ms, and a video processing speed of 125.6f/s, which can meet the real-time application requirements of micro unmanned aerial vehicles.