Chao Gao

Department of Computer Science and Engineering
Lehigh University
19 Memorial Drive W,
Bethlehem, PA, 18015
Email: Chg205 AT Lehigh DOT Edu

Towards Autonomous Wheelchair Systems in Urban Environments

(inspired by previous project ATRS)

In this work, we explore the use of synthesized landmark maps for absolute localization of a smart wheelchair system outdoors. In this paradigm, three-dimensional map data are acquired by an automobile equipped with high-precision inertial/GPS systems, in conjunction with light detection and ranging (LIDAR) systems, whose range measurements are subsequently registered to a global coordinate frame. The resulting map data are then synthesized a priori to identify robust, salient features for use as landmarks in localization. By leveraging such maps with landmark meta-data, robots possessing far lower cost sensor suites gain many of the benefits obtained from the higher fidelity sensors, but without the cost.

Data Acquisition

Our vehicle for data acquisition was "Little Ben," which previously had served as the Ben Franklin Racing Team's entry in the DARPA Urban Challenge. Vehicle 6-DoF pose was provided by an Oxford Technical Solutions RT- 3050. Range and bearing measurements from a pair of roof mounted, vertically scanning Sick LMS291-S14 LIDAR systems were then registered to the current vehicle pose to generate high-resolution range maps. The two LIDARs are highlighted (circled red) in the below figures. Two LIDARs were used to help reduce scene occlusion.

LIDAR Calibration

We placed retro-reflective targets on multiple poles and drove "Ben" around these poles to achieve sufficient point correspondences between the front and back LIDARs. These targets Can be automatically segmented using a threshold on the LIDAR remission measurements and geometry constraints. We employed a standard calibration approach to calibrate the front and back LIDARs. The calibration process consisted of two stages: (1) removing deterministic error between successive calibration loops caused by GPS "jumps". (2) removing the influences of random error in the calibration process.

Landmark Synthesis

We employed a recursive clustering approach in conjunction with several validation gates on cluster size and vertical level to segmente pole-like features. The below figures show the registered raw (left) and synthesized map data (right). The relative distance differences between synthesized landmark pairs and their real-world counterparts was about 7 cm.

Wheelchair Testing Platform

The smart-chair used in this work is based upon an Invacare M91 Pronto power wheelchair with Mk5 electronics. It integrates high resolution optical encoders, a Hokuyo UTM-30LX LIDAR system, a 1024X768 Point Grey digital video camera and a Garmin 18 WAAS enabled GPS system. For this work, the UTM-30LX was the wheelchair's primary sensor to detect pole-like features. In the current integration, the LIDAR and camera system are mounted on the opposite arm as the manual joystick controller as shown in the left figure.

Landmark Detection

Unlike the landmark synthesis phase, the wheelchair must rely entirely upon 2-D LIDAR scan data to identify 3-D pole-like features. We used clustering and circle fitting to segment pole-like feature candidates. Candidate features were then passed to the data association module to reduce false positives.

Localization Approach

We took EKF approach for estimating the 2-D wheelchair pose. In the prediction step, linear and angular velocities were estimated from the encoders. For the correction step, the observation functions were based upon LIDAR estimates for the range and bearing to the segmented landmark at position.

Landmark Position Refinement

To improve the inconsistency of the landmark map caused by GPS "jump", we introduced a refinement stage where SLAM is actually run by the wheelchair during an initial route traversal akin to a learning phase. For our implementation, we extended our EKF localization using a SLAM approach to further refine the landmark positions.

Experiment Results

Localization results using refined (right) and original landmark position estimates (left). Improving the consistency of landmark positions dramatically improved localization performance.

Videos

Publications