Building World Models for Mobile Platforms using Heterogeneous Sensors Fusion and Temporal Analysis

1 Introduction The Michigan Offroad Sensor Fusing Experimental Testbed (MOSlfEZ] is an autonomous vehicle designed to navigate an "own terrain which includes obstacles. lanes. and other traditional roadside obstructions. In order to gather data about the world around it, MOSFET is equipped with several heterogeneous sensors. These sensors include one forward-looking camera used for obstacle detection and lane tracking; two side-looking cameras used for lane sensing: and a bank of sonar sensors used to detect physical obstacles (see Figure 1). Figure 1: MOSFET MOSFEl's intelligence consists of three steps. First, it extracts information from each individual sensor. Next. it accomplishes a systematic fusion of the information extracted from the indwidual sensors. Finally, it augments this fused information by the use of temporal analysis, which keeps track of previous obstacle mformation and updates that information in accordance with the movement of the vehicle. MOSFET navigates by keeping within the lanes while simultaneously avoiding obstacles and other roadside obstructions. The crux of this paper is in steps 2 and 3, namely, the process of building a world model for MOSFET, based on the information gathered from the individual sensors: and augmenting this model through the use of temporal data. These 2 steps are accomplished in three stages: Deformable Template Stage-MOSFETs position and orientation. with respect to the left and right lane markers, are obtained by using STARLITE, a straight-edge detecting deformable template algorithm that uses a signal-to-noise ratio (SNR) matching criteria (see 111). The input to STARLITE consists of the left and right side-looking camera images. STARLITE's output is subsequently used to detect lane boundaries in the forward-looking camera image. Using position and orientation to fix three parameters of a parabolic deformable template, the lane's curvature is estimated by a Hough transform-like edge accumulation procedure. Correlation Stage-The ultrasound sensors are used to detect physical obstacles. As a separate step, a color segmentation algorithm called OLSON is also used to distinguish obstacles by analyzing the forward-lookmg camera image. By correlating the two obstacle extraction results, together with the lane information from the deformable template stage, the exact position and extent of obstacles that are relevant to MOSFETs navigability are determined The lanes and the obstacles detected are all laid out in a grid of MOSFETs immedate world. Each grid element represents a certain distance and angular position from MOSFET. Associated with each element is a value proportional to its "relative obstacleness." The relative …