2019 · Techmation · Harvey Mudd capstone
A senior capstone sponsored by Ningbo Techmation: an mmWave radar system that detects and tracks obstacles in real time for indoor autonomous guided vehicles (AGVs).
Cameras and lidar struggle in the conditions of a factory floor: dust, glare, low light, and reflective surfaces. Our sponsor, Ningbo Techmation, wanted to know how far a short-range mmWave radar could go as the only perception sensor on an indoor AGV. Could a robot navigate on its own and avoid obstacles, moving or still, using radar alone?
With no camera to fall back on, the tracking had to be accurate, and fast enough to run in real time on a small embedded computer.
Built a multi-object detection and tracking pipeline that ran live on an NVIDIA Jetson TX2. It used DBSCAN to cluster radar returns, gating and data association to follow each object between frames, and an Extended Kalman Filter to smooth and predict their motion.
Tested TI’s IWR1642 and IWR1443 radars against objects of different sizes and radar cross-sections, across a range of settings and indoor spaces, to learn what the sensor could and couldn’t see before trusting it for navigation.
Wrote a Python tool to parse the data packets streaming off the radar in real time. The rest of the pipeline depended on it.
Integrated the system on an indoor robot using ROS. With only the radar, an occupancy grid, and a cost map, the robot drove itself while avoiding static and moving obstacles.
The robot drove itself around an indoor course, using radar as its only sensor. I led the team in the spring semester and presented the work at the sponsor’s headquarters in Shanghai. We also measured the tracking system’s accuracy and latency, so the results were backed by numbers.