基于无人驾驶配送实验平台的封闭园区内动态环境视觉SLAM研究

Dynamic Environment Visual SLAM Based on Unmanned Delivery Experimental Platform in Closed Park Area

针对动态环境下SLAM算法定位信息精度不足、建图偏移严重问题,本研究基于无人驾驶配送实验平台提出一种面向封闭园区内动态环境的视觉SLAM系统。该系统以ORB_SLAM2为框架,添加了基于人体关键点提取的行为识别线程。在输入图像ORB特征点提取的同时,通过HRNet网络进行人体关键点的提取,通过30帧图像内容进行行为识别,判断图像中人体的行为状态是否为运动,并以此为依据筛选并剔除动态特征点,最后通过静态特征点进行位姿估计。实验结果表明,在TUM数据集动态子序列下,与ORB_SLAM2、DS_SLAM相比,本研究所提算法在系统精度与速度上达到平衡,有效提高了位姿估计的准确性。

Our research proposes a visual SLAM system tailored to dynamic environments within enclosed park area,aiming to address the issues of low localization accuracy and severe map offset in SLAM algorithms.Based on an unmanned delivery experiment platform,this system utilizes ORB_SLAM2 as the framework,and supplemented with a behavior recognition thread based on human keypoint extraction.While extracting ORB feature points from input images,the system also employs HRNet network to extract human keypoints,conducts behavior recognition based on 30-frame image content to determine the movement status of human subjects in the images,and subsequently screens and eliminates dynamic feature points accordingly.Finally,pose estimation is carried out using static feature points.Experimental results demonstrate that,compared with ORB_SLAM2 and DS_SLAM,our research proposed algorithm achieves a balance between system accuracy and speed in dynamic subsequences of the TUM dataset,effectively enhancing the accuracy of pose estimation.