Tactic-Based Motion Modelling and Multi-Sensor Tracking

Tracking in essence consists of using sensory information combined with a motion model to estimate the position of a moving object. Tracking efficiency completely depends on the accuracy of the motion model and of the sensory information. For a vision sensor like a camera, the estimation is translated into a command to guide the camera where to look. In this paper, we contribute a method to achieve efficient tracking through using a tactic-based motion model, combined vision and infrared sensory information. We use a supervised learning technique to map the state being tracked to the commands that lead the camera to consistently track the object. We present the probabilistic algorithms in detail and present empirical results both in simulation experiment and from their effective execution in a Segway RMP robot. Introduction There have been a number of investigations into the problem of tracking moving objects e.g. (Doucet, Freitas, & Gordon 2001). Within the robotics community, there has been a similar interest in tracking objects from robot platforms e.g. (Schulz, Burgrad, & Fox 2003). When tracking is performed by a robot executing specific tasks acting over the object being tracked, such as a Segway RMP soccer robot grabbing and kicking a ball, the motion model of the object becomes complex, and dependent on the robot’s actions (Kwok & Fox 2004). In this paper we show how multiple motion models can be used as a function of the robot’s tactic using a particle-filter based tracker. Over the years, a lot of different sensors such as vision sensors, infrared and ultrasound sensors have been used in the robotics community. For environments the Segway RMP operates in, there are few sensors that can compete with color vision for low cost, compact size, high information volume and throughput, relatively low latency, and promising usage for object recognition (Browning, Xu, & Veloso 2004). Thus, we choose vision as the primary sensor. Recently, we have equipped each robot with a infrared sensor to reliably detect objects close to it. We introduce how this additional information can be of use for tracking. Copyright c 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. In order to fully utilize the tracking information, the state being tracked has to be translated to the control command to guide the camera where to look. Our previous implementation of this translation is completely based on the geometric model. We recently have used a supervised learning technique to do the mapping. The paper is organized as follows. In the following section we give a brief description of the Segway RMP soccer robot. Next we describe the tactic-based motion modelling. We show the multi-sensor multi-model tracking algorithm. We then focus on our supervised learning technique on camera control, leading to our experimental results, related work, conclusions and future work. Segway RMP Soccer Robot The Segway platform is unique due to its combination of wheel actuators and dynamic balancing. Segway RMP, or Robot Mobility Platform, provides an extensible control platform for robotics research (Searock, Browning, & Veloso 2004). In our previous work, we have developed a Segway RMP robot base capable of playing Segway soccer. We briefly describe the two major components of the control architecture, the sensor and robot cognition, which are highly related to our tactic-based motion modelling for efficient tracking. Vision Sensor and Infrared Sensor The goal of vision is to provide as many valid estimates of objects as possible. Tracking then fuses this information to track the most interesting objects (a ball, in this paper) of relevance to the robot. We do not discuss the localization of the robot in the sense that a lot of soccer tasks (known as tactics in later sections) can be done by the Segway RMP robot independently of knowing where it is in the world. Also we use global reference in this paper (global position and velocity) which means it is relative to the reference point where the robot starts to do dead reckoning. The infrared sensor is added to detect the ball when it is in the catchable area of the robot. Its measurement is a binary value indicating whether or not the ball is in that area. In most cases, this is the blind area of the vision sensor. Therefore, the infrared sensor is particularly useful when the robot is grabbing the ball. Furthermore, it works very reliably so that we assume its measurement “is” the true value. This