Overview

All the source codes are in the ugv_stereo GitLab workspace. They are tagged with final version. The core modules include:

• SSLAM package is an optimization-based multi-sensor (mainly stereo camera) state estimator, which achieves accurate self-localization and map management for autonomous applications.

• Cubicle Detection package handles stereo matching for range estimation, obstacle detection, object classification and tracking, using Darknet as backend.

• LaneKerb Detection package achieves real-time lane and kerb detection under various road scenarios and diverse weather conditions.

• Cubicle Merge package handles the mapping functionality. It merges the object tracking output with localization 6D pose output.

• Calibration package convert the Matlab calibration in to the desired format for this application.

Platform

Ubuntu 16.04 or above. Fully tested under Ubuntu 16.04.

ROS envrionment.

System Diagram

Fig. 1 System Flow Chart.

Executive Summary of different modules

Module	Theory	Dependencies	Latency (ms)
SSLAM	Optical Flow Good Feature to track Sliding window optimization	OpenCV Ceres-solver Cereal	100
Cubicle Detection	Stereo matching Postive obstacle detection Negative obstacle detection Multiple object classification Multiple object tracking Slope estimation	OpenCV Darknet with CUDA	40
Curb & Lane Detection	Fully Convolutional Networks Mean Shift clustering	Tensorflow 1.10	30
Cubicle Merge			30