基于深度学习的智能叉车障碍物识别与定位算法  

Intelligent obstacle identification and location algorithm of forklift based on deep learning

作　　者：赵钢 曹男 曹轶伦 Zhao Gang;Cao Nan;Cao Yilun

机构地区：[1]中铁联合国际集装箱智慧物流成都有限公司,成都610084 [2]西南交通大学机械工程学院,成都610031

出　　处：《起重运输机械》2025年第6期34-40,共7页Hoisting and Conveying Machinery

基　　金：国家自然科学基金(51675450)。

摘　　要：实现叉车障碍物识别与定位对提升物流仓储行业自动化、智能化具有重要的意义。现有技术大多受限于实际生产中的复杂场景,无法满足对障碍物识别与定位的高精度实时性的要求。针对现有问题,文中提出了一种基于深度学习的智能叉车障碍物识别与定位方法,以实现在仓储货场环境下对障碍物实时、高精度的识别与定位。首先,设计了改进的YOLOX-nano网络获取障碍物外接矩形框,实现对障碍物的检测与提取;然后设计了基于改进的DeepLabv3+网络对筛选出的障碍物进行分割;最后采用阈值化分割方法对障碍物分割区域进行轮廓提取,进而求解障碍物的像素中心点。在叉车实际工作环境中的障碍物数据集上进行实验,结果表明,该方法能够快速准确地对障碍物物中心进行定位。其中第1阶段的障碍物检测平均精度均值为99.90%,平均检测时间为25.62 ms;第2阶段障碍物分割的平均像素准确度为95.93%,平均分割时间为24.69 ms,保证叉车对障碍物实时识别的同时具有较高的准确率。Accurate identification and location of obstacles by forklifts are crucial for enhancing the automation and intelligence of the logistics and warehousing industry.However,due to the complex scenarios encountered in production,most existing technologies fail to meet the stringent requirements for high-precision,real-time obstacle detection and localization.To address these challenges,an intelligent forklift obstacle identification and location method based on deep learning is proposed.This method aims to achieve real-time,high-precision obstacle identification and localization within the warehouse yard environment.Initially,an enhanced YOLOX-nano network was designed to identify and extract obstacles by obtaining their bounding rectangles.Subsequently,an improved DeepLabV3+network was employed to segment the detected obstacles.Finally,a threshold segmentation method was adopted to extract the contours of the segmented obstacle areas,and the pixel center of each obstacle was calculated.Experiments conducted using an obstacle dataset collected from the actual working environment of forklifts demonstrate that this method can swiftly and accurately locate the centers of obstacles.In particular,the first stage of obstacle detection achieved an average accuracy of 99.90%,with an average detection time of 25.62 ms.In the second stage,the average pixel accuracy of obstacle segmentation reached 95.93%,and the average segmentation time was 24.69 ms.These results demonstrate that this method can ensure real-time obstacle identification with high accuracy,thereby significantly enhancing the safety and efficiency of forklift operations.

关 键 词：智能叉车 目标检测 目标分割 障碍物识别 目标定位 

分 类 号：TH242[机械工程—机械制造及自动化]