英文摘要 | Person following is an important ability for mobile robots, which can be applied for carrying heavy loads and offering immediate services by following a specified person. It is significant in both theoretical research and real-world applications. This thesis focuses on target locating and collision-free following for person-following robots. The contents are as follows:
Firstly, the research background and its significance of person following for mobile robots are given. The research development of target person locating, person-following approaches, and distraction-handling schemes are reviewed. The contents and structure of this thesis are also briefed.
Secondly, a visual T-D-R target person locating approach is proposed. The target person is located with a correlation filter-based visual tracker in a short time, and the problems of determining tracking quality and tracking recovery are solved by the combination of a pedestrian detector and a person re-identification module. The re-identification module is comprised of a CCF module and an IDLA module, where the former online learns convolutional features and a discriminative correlation filter with a linear kernel, and the latter stores several recent images of the target person and identifies the target person with an end-to-end person re-identification network. Combining the stable identification ability of the CCF module with the adaption to appearance variations of the IDLA module, the re-identification module can satisfy the requirement of long-term and robust tracking. The proposed approach can effectively distinguish target person in multi-pedestrian environments, which is verified by the comparisons with state-of-the-art approaches as well as the person-following experiment on the quadruped robot SCalf-III.
Thirdly, a target person locating approach based on a pedestrian detector and data association is presented, where the pedestrian detector is used to provide candidate locations of the target person, and a Kalman filter and a person re-identification module are employed to identify target person by associating detection results. The Kalman filter with a motion model is applied to predict the bounding box of the target person, which is then used to compute spatial relationships with detection results. The person re-identification module is introduced to improve the reliability of data association in multi-pedestrian environments, where the Siamese network-based SN module and IDLA module are integrated to measure the similarity between detection results and the target person. On this basis, an effective segmentation algorithm of LiDAR point cloud is given, which can be used to obtain non-ground object clusters. 3D pedestrian locating and clusters classification can be implemented by integrating visual pedestrian semantic information with non-ground clusters. The method is verified by the experiments.
Fourthly, a collision-free following approach based on path planning is proposed. The static obstacle cost map is generated according to the results of the point cloud segmentation algorithm. Multiple disturbing pedestrians are tracked based on unscented Kalman filter (UKF) and nearest near joint probabilistic data association (NNJPDA), and their motion states are estimated continuously by matching 3D pedestrian locations and tracking trajectories. Then, the dynamic pedestrian cost map is generated. On this basis, the global planner plans a global path with the static obstacle cost map, whereas the local planner also combines the dynamic pedestrian cost map for an optimized local path. This local path tends to follow the global path and avoid the collision with dynamic pedestrians. By combining low-frequency global planning with high-frequency local planning, the mobile robot achieves a safe and collision-free person following, which is verified in both indoor and outdoor environments.
Fifthly, aiming at the problem where the target person is beyond the field of view of robot sensors for a long time, a target person locating framework based on distributed odometries is proposed. The robot-end odometry and target-end odometry are used to estimate poses of the robot and positions of the target person, respectively, which is then integrated with 3D locations of the target person in robot coordinate to calculate the coordinate transformation between these two odometries. On this basis, the mobile robot can obtain the estimated target location by receiving the position information from the target-end odometry, which is then transformed to the coordinate in the robot-end odometry. When the mobile robot cannot locate the target person with its own sensors, it strives to re-detect the target person by moving along the sampling transformed positions using the collision-free following approach. With the robot-end LiDAR odometry and target-end wearable visual-inertial odometry, experiments are conducted and the results show that a mobile robot can find target person and recover person following even if the target person is far away from the robot.
Finally, conclusions and future work are offered. |
修改评论