基于视觉的行人引领移动机器人导航方法研究

移动机器人通过跟随领航员以实现导航是一种便捷的导航方式。针对行人引领导航中的领航员定位问题,提出了一种基于视觉的行人引领导航方法。该方法利用卡尔曼滤波器预测领航员的位置和尺度,并基于深度神经网络的行人检测器提供的结果更新滤波器的状态。为了关联检测结果和卡尔曼滤波器预测结果,提出了2个指标用于衡量两者之间的关联性。其中,为了提高在多个行人中辨认领航员的可靠性,创新性地引入了一个孪生神经网络,使用该网络全连接层提取的特征作为候选人的特征描述子,并通过计算特征之间的余弦距离来验证检测器检测到的行人身份。此外,当卡尔曼滤波器跟踪领航员失败时,综合考虑检测结果和孪生网络的判断结果重新初始化卡尔曼滤波器,以实现持续的领航员定位。视频实验和物理机器人实验验证了所提出的方法的有效性和可靠性。

A Vision-based Leader-Guided Navigation Method to Mobile Robots

Navigation by following a specified leader is a convenient way for mobile robots. In order to solve the problem of the leader positioning, a vision-based leader-guided navigation approach for mobile robots is proposed. A Kalman filter is utilized to predict the locations and scales of the leader, which updates the states according to the results obtained by a CNN-based pedestrian detector. Two criteria are used to evaluate the relationship between the detection results with the result of the Kalman filter. Specifically, a Siamese network is employed to improve the reliability of distinguishing the leader among multiple pedestrians. The features extracted from the fully connected layer of the Siamese network are innovatively adopted to verify the identities of the pedestrians detected by the detector through calculating the cosine distances between these features. In addition, when the Kalman filter fails to track the leader, the Kalman filter is re-initialized by considering the detection results and the judgment results of the Siamese network, so as to achieve continuous navigator positioning. The effectiveness and reliability of the proposed navigation approach is verified by video experiments and actual experiments on a mobile robot.