基于人类行为下的多目标跟踪视觉时序注意控制算法研究(英文)

Human behavior inspired temporal attention control system for multiple object tracking

由于移动式机器人所携带的传感器资源有限,能够在时间维度上共享传感器资源,实现多目标跟踪,是智能化机器人重要的研究领域之一.本文提出了一个基于人类行为模型的时序注意控制算法,能够实现使用单个2自由度摄像头实时监控多个自由移动的目标物体.该算法基于以下三个控制原则:①最小化监控目标位置预测的不确定性以及各个目标物预测效果的差异;②最小化多目标跟踪过程中摄像头切换视角所需的能量消耗;③最大化能够呈现在摄像头视觉范围内的目标物体的个数.算法根据当前检测到目标信息,通过卡尔曼滤波,预测目标物的运动规律,选择下一时刻摄像头的最佳注意视角.该算法通过matlab仿真以及在移动式机器人实体上进行了实验验证,结果显示,该算法能够在时间顺序有效的共享单个摄像头资源,检测摄像头视角范围内的目标的位置,并根据上一个时间点的检测信息预测视角范围外的目标,实现跟踪多个目标物体,降低预测及跟踪的误差,并优化跟踪过程中切换摄像头视角的能量损耗.

An important ability for mobile robots is to process multiple tasks in complex environments. Since the sensor resources on a robot are limited, it is necessary to distribute the sensors＇ attention to different tasks along the time scale. A human inspired temporal attention control method is proposed, which aims at detecting multiple objects and estimating their poses with a single actuated camera. The proposed method is based on three criteria which are partially inspired by human behavior： （1） Minimization of the overall object poses perception uncertainty and minimization of the variance of the perception uncertainty of different objects; （2） Minimizing the energy cost of the camera movements for completing the tasks; （3） Maximization of the number of objects in the camera＇s field of view. Kalman filters are used to estimate the object poses and to detemaine the perception uncertainty. The proposed approach is tested in both simulation and an experiment on the robot. The results show that the proposed approach is able to switch the camera＇s attention according to the object＇s states efficiently with a low frequency of camera movements.