Algorithms for LiDAR Based Traffic Tracking: Development and Demonstration
The current state of the art of traffic tracking is based on the use of video, and requires extensive manual intervention for it to work, including hours of painstaking human examination of videos frame by frame which also make the acquisition of data extremely expensive. Fundamentally, this is because we do not have observability of the actual scene from a camera which captures a 2D projection of the 3D world. Even if video were to be automated, it would involve such algorithms as RANSACK for outlier elimination while matching features across frames or across multiple cameras. This results in algorithms without stationary relationships between input and output statistics, i.e., between sensing resolution and error and estimated positions and velocities. LiDAR directly provides 3D point clouds, giving a one-one mapping between the scene from the physical world and data. However, available eye-safe lidars have been developed for autonomous vehicles, and provide only sparse point clouds when used for longer range data acquisition.^ Our experimental results use the Velodyne HDL 64E lidar. The sparse nature of data points returned by the Velodyne LiDAR rendered most of the algorithms for object identification and tracking using 3D point clouds at the point cloud library (PCL), a leading multi-agency open source research initiative focused on 3D point cloud processing ineffective for our work. Hence I developed a comprehensive set of algorithms developed to identify and remove background; detect objects through clustering of remaining points; associate detected objects across frames, track the detected objects, and estimate the dimension of objects.^ Two different complementary algorithms based on, surface equation (in 3D Cartesian coordinates) and LiDAR spherical coordinates were developed for background identification and removal. Delaunay triangulation based clustering is performed to identify objects. Kalman filter and Hungarian assignment algorithm are used in tandem to track multiple objects simultaneously. A novel bounding box algorithm was devised taking advantage of the way LiDAR scans the environment to predict the orientation and estimate dimension of objects. Trajectory analysis is performed to identify and split any wrong associations, join trajectories belonging to same object and stitch partial trajectories. Finally, the results are stored in a format usable by various transportation or traffic engineering applications.^ The algorithms were tested by peers with data collected at three intersections. Detection rate and counting accuracy are above 95% which is on par with commercial video solutions that employ humans to varying degrees. While prototyping for the algorithms was done it MATLAB, preliminary tests of conversion to C++ showed that the developed algorithms can be executed in real time on standard computer hardware.^
Kartik B. Ariyur, Purdue University, Andew P. Tarko, Purdue University.
Civil engineering|Mechanical engineering|Robotics