The Mapping Trike Project
Among the many challenges outdoor navigation poses, robust localization is one of the most difficult problems to solve. Until recently, most solutions rely on SLAM algorithms because of a widely held notion that outdoor environments are simply too large to map a priori. However, this conclusion has been proven incorrect time again as many large companies, such as Google, routinely deploy custom mapping vehicles that can efficiently gather vast amounts of data in even the most remote locations. The smart wheelchair project is designed with this idea in mind. It uses its visual sensing suite in order to detect potential landmarks, which are compared to a pre-computed map of landmarks it carries. The Lehigh Mapping Trike (LMT) is motivated by the lack of means to construct the necessary large-scale maps in outdoor pedestrian zones.
The LMT features 3 Sick LMS219 LIDARs, a GPS module, an inertial measurement unit (IMU), an encoder, and an onboard computer. The system is powered with two 12V/12Ah batteries, providing about two hours of continuous run time.
Two of the three LIDARs are rotated to take vertical scans off of the port and starboard side of the LMT. As the LMT moves forward the vertical scans can be aggregated to construct a large 3D reconstruction of the space captured. While these 3D clouds can provide much detail for more accurate landmark detection, they can only observe a single landmark once. Therefore, a third LIDAR is mounted normally on the stern taking planar scans. As a result, a single landmark can be seen multiple times over a period; the trade off is less information.
Data from the GPS module, IMU, and encoder, are fused together in order to produce an accurate estimate of position and orientation. It is important to note that the GPS provides a coarse estimate of position, and is mainly used to define the location of landmarks with respect to coordinates on a UTM plane.
Map generation occurs off-line and involves four stages. The first step involves computing accurate 3D pose estimates at every time step of the data collection. The second stage synthesizes sliding windows of point clouds from the separate laser scans and removes the ground plane. Next, poles are detected in the remaining cloud (poles such as street signs, lamp posts, and parking meters are landmarks in the map). Finally, a SLAM algorithm is used to optimize the location of landmarks relative to the path the LMT took.
An Extended Kalman Filter (EKF) is used to fuse data from the GPS, IMU, and encoder, in order to estimate a 3D position and orienation. The final state is a five dimensional vector containing X, Y, Z, yaw, and pitch; roll is ignored. Below is a visualization of odometry data gathered from the LMT's path around the Packer Ave, Vine St, Morton St, and Webster St, block in South Bethlehem.
Separate pole segmentation algorithms are used to detect landmarks in point cloud data gathered from the side LIDARs and the stern LIDAR. While detection with the stern LIDAR is more straight forward, detection with the side LIDARs requires more thought.
First, each individual vertical scan, taken with respect to the robot frame, must be transformed to the world frame and aggregated. A sliding window effect is created by maintaining a certain number of scans at a given time step. The ground plane is removed from the resulting cloud and the remaining points are clustered, leaving only the most dense set of points in the cloud. Most of what remains are usually poles, although other artifacts such as cars, people, and buildings also persist. Finally, based on certain criteria the process determines which of the remaining clusters are poles. Each pole is tied to the odometry step at which it is first seen.
The last stage of the software architecture consists of generating a 2D map of landmarks by optimizing the location of observations over the path the LMT took. Interest in 2D maps stems from the use of a "flat-earth" model for the Smart Wheelchair Project. As a result, the LMT seeks to define the location of landmarks with respect to the UTM plane the LMT is currently navigating on. Currently, the efficacy of a few different SLAM algorithms are being investigated. A map from the first iteration of the system can be seen below, which was generated using EKF SLAM with the data from the trip around the the Packer Ave, Vine St, Morton St, and Webster St, block in South Bethlehem.
Hallway Mapping using the Trike
- C. Gao and J. R. Spletzer, "On-line Calibration of Multiple
LIDARs on a Mobile Vehicle Platform," the 2010 IEEE International
Conference on Robotics and Automation, Anchorage, Alaska, May
Full Text (PDF) - bibtex
- Kyle Hart
- Tashwin Khurana