改进的SlimYOLOv3-A算法在障碍物检测中的应用  

Application of Improved SlimYOLOv3-A Algorithm in Obstacle Detection

作　　者：汪刚 魏赟[1] 张宇辰 WANG Gang;WEI Yun;ZHANG Yuchen(School of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《计算机与数字工程》2023年第9期2019-2025,共7页Computer & Digital Engineering

基　　金：国家重点研发计划项目(编号:2018YFB1700902)资助。

摘　　要：针对移动物流机器人的工作环境光线暗、背景复杂,障碍物目标小、特征不显著等问题,考虑到物流机器人处理器设备的内存和计算能力有限,提出一种在卷积层进行通道剪枝的高效目标检测算法。对YOLOv3目标检测算法进行了改进,提出了训练参数和运算量都相对较少的SlimYOLOv3-A:在YOLOv3的backbone尾端添加PSA注意力模块,提高对小目标障碍物的检测精度;在通道比例因子上施加L1正则化来改善卷积层的通道稀疏性,修剪信息量较少的特征通道以获得精简的目标检测模型。实验结果表明,改进后的模型有效地平衡了模型大小、检测性能和运算速度三者的关系。Slim YOLOv3-A(90%剪枝率)对比YOLOv3的参数量减少约87.5%,模型体积减少约87.4%,检测速度提高43%,检测精度提升2.6%。Aiming at the problems of the mobile logistics robot's working environment,such as dark light,complex background,small obstacle and fuzzy features,and considering the limited memory and computing power of the logistics robot's processor,an efficient target detection algorithm based on channel pruning in the convolutional layer is proposed.A target detection algorithm based on YOLOV3 is improved,SlimYOLOV3-A,with few training parameters and workload.PSA attention module is added to backbone of YOLOv3 to improve the detection accuracy of small target obstacle.L1 regularization is applied to the channel scale factor to improve the channel sparsity of the convolutional layer,and feature channels with less information are pruned to obtain a streamlined target detection model.The experimental results show that the model can effectively balances between model size,detection performance and computing speed.Compared with YOLOv3,Slim YOLOv3-A(90% pruning rate)reduced the number of parameters by about 87.5%,model volume by about 87.4%,and detection speed by 43%,and the detection accuracy is improved about 2.6%.

关 键 词：障碍物检测 稀疏训练 通道修剪 注意力模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]