Perception GPU Obstacle Detector Module

About

This CUDA-based module performs real time obstacle detection for the rover and sends LCM messages containing information about obstacles. Requires an NVIDIA GPU and CUDA installation to run.

Code walkthrough: https://youtube.com/playlist?list=PLj5rOuXit19mCtakn21m11At3zjgEycnO

Dev Enviornment Setup

Follow the install guide for your desired workflow.

• Local: https://docs.google.com/document/d/1MELSAlFY7PWr0ZisBeJtXIB8x-YFIi4BrLiyycQkH5Y/edit
• GreatLakes: https://docs.google.com/document/d/1Xkk615z5T1gS0-P0z-_xr7CUilrxN_Vcua-Uol290bg/edit?usp=sharing

Usage

Great Lakes:

1. Connect to UMich VPN (if off campus) and log into: https://greatlakes-oncampus.arc-ts.umich.edu/pun/sys/dashboard
2. Go to My Interactive Sessions and launch a new session on GPU partition with 1 GPU
3. Run ./launchme to open the Singularity container for development
4. Run source sourceme to add CUDA compiler to your shell
5. From mrover-workspace do:

./jarvis build jetson/percep_obs_detect
./jarvis exec percep_obs_detect

Local:

1. From mrover-workspace do:

./jarvis build jetson/percep_obs_detect
./jarvis exec percep_obs_detect

Overview

obs-detector: Driver program, operating mode selection

common: Lazy utility file

pass-through: Simple filter to remove points too close or too far in our view

plane-ransac: RANSAC Plane segmentation for ground detection and removal

euclidean-cluster: Detects obstacle bounding boxes after all filtering

path-finder: Solves a clear path based on bounding box detections

Other

Deprecated documentation: https://docs.google.com/document/d/1WPW7yxcCp_EcDjPbGVOtrZl3A5krSXDas_pcxD3N2IM/edit