基于距离选通图像的目标检测方法研究  

Research on Object Detection Method Based on Distance Range-Gated Images

作　　者：田青[1] 王述敏 温博 

机构地区：[1]北方工业大学信息学院,北京100144 [2]北京师范大学-香港浸会大学联合国际学院,广东珠海519087

出　　处：《工业控制计算机》2024年第12期28-30,33,共4页Industrial Control Computer

基　　金：国家重点研发计划资助(2020YFB1600702)。

摘　　要：由于距离选通图像缺少颜色特征信息导致检测率低、距离信息特征提取能力较弱等问题,因此,提出了一种基于改进YOLOv8s网络的距离选通图像检测方法。首先,提出自适应选通特征融合模块,对距离选通图像和其深度信息进行特征级融合,获取高级语义信息;其次,采用基于多尺度的深度可分离卷积模块构成的空间注意力模块,对每个特征层进行加权运算,以获得更多的关键特征;最后,改进了Neck部分PAN结构,提高了网络的目标检测能力。在距离选通图像数据集上进行了实验,实验结果表明该方法能有效地提升检测精度,并具有实际应用价值。Since the lack of color feature information in range-gated images leads to low detection rate and weak feature extraction of distance information,a range-gated images detection method based on improved YOLOv8s network is proposed.Firstly,an adaptive range-gated feature fusion module is proposed to perform feature-level fusion of the range-gated image and its depth information to obtain high-level semantic information.Secondly,a spatial attention module based on the composition of multiscale depthwise separable convolutions modules is used to perform a weighting operation on each feature layer to obtain more key features.Finally,the Neck part of the PAN structure is improved to enhance the network's target detection ability.Experiments are carried out on the range-gated image dataset.

关 键 词：距离选通图像 深度学习 目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]