Abstract: Mobile robots have broad market application prospects in modern integrated intelligent manufacturing due to their advantages such as reliable performance and reduced manufacturing costs.However,there exist problems such as single perception dimension,low precision in the indoor environment perception system of mobile robots,which is hard to meet the demands of multi-dimensional real-time and accurate perception of the environment at present.Therefore,an environment perception system that could simultaneously perform 2D dynamic target detection and 3D scene map reconstruction in a laboratory environment was constructed by simulating a logistics warehouse,the images collected by the binocular camera were respectively input into the improved You Only Look Once version 3(YOLOv3)network fused with Tree-structured Feature Aggregation(TFA)module and the optimized RTAB-MAP algorithm with Multi-Scale Retinex with Color Restoration(MSRCR)algorithm and key frame filtering mechanism,thus the results were displayed in Robot Visualization tool(RVIZ)in real time and achieved multi-dimensional environment perception,so that it could meet the needs of multi-tasking.The experimental results showed that the precision and recall in dynamic target detection were improved by 1.78% and 1.73% respectively compared with the original algorithm,the detection time was 16.57ms/f,and the average positioning error was 1.49%.Compared with the original algorithm,the errors of the improved RTAB-MAP algorithm were significantly reduced,and the quality of the 3D point cloud map in the actual indoor scene reconstruction was better.

Key words: binocular vision, moving target detection, 3D reconstruction, YOLOv3, mobile robot

摘要： 移动机器人凭借其工作性能可靠、减少生产制造成本等优势,在现代集成智能制造业中具有广阔的市场应用前景。但目前移动机器人室内环境感知系统存在感知维度单一、精度不高等问题,往往难以满足多维度实时精确感知环境的需求。因此,在模拟物流储仓的实验室环境下构建了一种可同时进行二维动态目标检测和三维场景地图重建的环境感知系统,将双目相机采集的图像分别输入增加树形特征融合模块的改进YOLOv3网络与加入带色彩恢复的多尺度视网膜增强算法和关键帧筛选机制的优化RTAB-MAP算法,运行结果在机器人可视化平台实时显示,从而进行多维度环境感知,满足多任务需求。实验结果表明,动态目标检测中查准率与查全率较原算法分别提高1.78%和1.73%,检测耗时为16.57ms/f,平均定位误差为1.49%;改进后的RTAB-MAP算法相较原算法各误差均显著下降,实际室内场景重建中三维点云地图质量更佳。

关键词: 双目视觉, 运动目标检测, 三维重建, YOLOv3, 移动机器人