面向嵌入式系统的轻量级目标检测算法

Lightweight target detection algorithm for embedded system

由于嵌入式设备的内存和计算能力有限,在无人机平台上运行基于深度学习的目标检测算法进行实时解析具有很大的挑战性,同时,算法在小目标、高密度、多类别的场景下,检测精度有待提高。在此前提下,构建NCWS-YOLO轻量级算法,以YOLOv5算法为基础,基于非极大值抑制方法,融合加权框融合算法,提出了NCW方法,重构了预测端目标框筛选网络,使检测精度提升了3.9%。并且利用通道剪枝技术,对批归一化层进行通道稀疏化训练,选择不包括shortcut的层进行修剪,使参数量减少了74%,模型大小缩减了72.2%,浮点数运算降低了37.6%,将算法部署于嵌入式设备上实现了对无人机数据集的目标检测任务。所提方法在无人机数据集上测试精度(Pr)和平均精度(mAP@0.5)分别达到了0.941和0.969,在Nvidia Jetson TX2上推理速度提升了49.6%。实验数据表明,该网络能够在低功耗、算力低的嵌入式设备上进行实时检测。

Due to the limited memory and computing power of embedded devices,it is very challenging to run the target detection algorithm based on deep learning on the UAV platform for real⁃time analysis.At the same time,the detection accuracy of the algorithm in small target,high density and multi⁃category scenarios needs to be improved.Under this premise,constructing the NCWS-YOLO lightweight algorithm,based on the YOLOv5 algorithm,by combining weighted frame fusion and non⁃maximum suppression for the classification target,NCW method is proposed,and the target frame screening network at the prediction end is reconstructed,which improves the detection accuracy by 3.9%.And using the channel pruning technology,training the batch normalization layer for channel sparseness,selecting the channel that excluding layer of the shortcut to prune,the amount of the parameter was reduced by 74%,the size of the model was reduced by 72.2%,the floating point operation was reduced by 37.6%,deploying the algorithm on embedded devices,and the target detection task of UAV data set was realized.The proposed method tested Precision(Pr)and mean Average Precision(mAP@0.5)on the UAV data⁃set was reached 0.941 and 0.969,and the inference speed increased by 49.6%on Nvidia Jetson TX2.Experimental data shows that the network can perform real⁃time detection on embedded devices with low power consumption and low computing power.