实时的移动机器人语义地图构建系统

发布时间：2018-07-31 09:59

【摘要】：语义信息可以使机器人更充分地理解未知环境,为更高级的人机交互和完成更复杂的任务奠定基础。为了能够使移动机器人实时地创建语义地图,在Jetson TX1嵌入式电脑上开发了一种轻量级的深度学习目标检测模型,在保证了检测精度的同时,实现了高效的目标检测功能。并利用了视频流中的帧间光流信息,使用运动信息指导传播算法降低检测算法的漏检率。对于Kinect传感器生成的深度图像有黑边、黑洞等缺陷,使用统一计算设备架构(CUDA)技术开发了一种实时的深度图像修复算法。利用即时定位与地图构建(SLAM)技术,实现移动机器人底层的定位、导航、地图创建功能,并在此基础上使用贝叶斯推理框架,同时融合了环境的度量信息与视觉识别信息完成了语义地图的创建。经过实验表明,所提出的方法在实际的、复杂的室内环境下可以使移动机器人实时地创建语义地图。
[Abstract]:Semantic information can make the robot fully understand the unknown environment and lay the foundation for more advanced human-computer interaction and the completion of more complex tasks. In order to make the mobile robot create semantic map in real time, a lightweight depth learning target detection model is developed on Jetson TX1 embedded computer, which ensures the detection accuracy and realizes the efficient target detection function. Using the inter-frame optical flow information in the video stream, the motion information guidance propagation algorithm is used to reduce the detection rate of the detection algorithm. For the defects of the depth images generated by Kinect sensors such as black edges and black holes, a real-time depth image restoration algorithm is developed using unified computing equipment architecture (CUDA) technology. Based on the (SLAM) technology of instant location and map construction, the functions of location, navigation and map creation in the bottom layer of mobile robot are realized, and the Bayesian reasoning framework is used on this basis. At the same time, it combines the measurement information of environment and visual recognition information to complete the creation of semantic map. The experiments show that the proposed method can make the mobile robot create semantic map in real time and in complex indoor environment.
【作者单位】：北京工业大学信息学部;数字社区教育部工程研究中心;
【基金】：北京市教育委员会科技计划面上项目(KM201510005005) 北京工业大学智能制造领域大科研推进计划(040000546317552)项目资助
【分类号】：TP242;TP391.41

【相似文献】