| With the development of robot technology,all kinds of patrol robots and service robots get closer to our daily life.Environment perception and localization becomes more important for robots as they coexist with pedestrians.Environment perception enables robots to understand the environment around them.Based on the understanding like where human is and where they can pass through,robots can do some more complicated work.Meanwhile,robots need to know where themselves are.Robots are able to move around in the environment only when they can locate themselves.To solve these two problems,a new framework for environment perception and simultaneous localization and semantic mapping is proposed in this dissertation.It's assumed that traversable region is region where human can get through.Based on the assumption,pedestrians in the environment are detected effectively and a small area near the feet of pedestrians is sampled as traversable region.Then,a Multivariate Gaussain model is built for the region and the region growing algorithm is applied to detect all the traversable regions in the image based on the Gaussain model.The experiments shows that the method performs well in many complex environments.A framework like ORB-SLAM is adopted in the simultaneously localization and semantic mapping part.The whole algorithm in this part is divided into three threads: tracking,local mapping,and loop closing.In tracking thread,a constant velocity+DLT(Direct Linear Transformation)and greedy search+DLT algorithm is proposed to match features extracted at different poses effectively.It's proved that the algorithm is more robust and produces more accurate matching results than the algorithm adopted in ORB-SLAM.In local mapping thread,an weighting strategy based on the quality of map points is proposed to optimize robot poses and map points.And the travesable region detected previously is projected into world coordinate to produce a semantic map with traversable region information of the environment.When tested in real environment,our algorithm turns out to be more robust than ORB-SLAM,and it can output a semantic map with traversable region information in the environment. |
