Low cost stereovision based obstacle detection and speed control of autonomous buggy

In recent years, Singapore is witnessing autonomous vehicle trials in public places as an initiative of “Smart Mobility 2030”. One such attempt is Energy Research Institute @NTU’s (ERI@N’s) attempt to fully automate a manually driven buggy, which is currently deployed in Bus Interchange of Woodlands MRT. Conversion of a manually driven vehicle into a driverless vehicle involves several changes or additions in software and hardware for localization, mapping, perception, tracking, sensor fusion etc. This thesis focusses on design, development and testing of a low cost stereovision based obstacle detection based on which buggy’s speed is controlled autonomously. The project is divided into two phases. The first phase includes the development of vision algorithm to detect and filter the obstacles in the buggy’s path. The second phase involves the retrofitting additional hardware to the vehicle’s speed control system to autonomously control its speed to avoid collision with detected obstacles. The proposed vision system consists of a low cost stereoscopic camera for image acquisition. As a dedicated buggy lane is provided and since people are the main obstacles in the path, a robust human detection algorithm, Histogram of Oriented Gradients (HOG) based feature extraction with Support Vector Machine (SVM) based classification, is used. After the detection, further processing is done on the image to obtain the distances of the obstacles from vehicle by exploiting the stereoscopic property of the camera. In the second phase of the project, the existing speed control system of the buggy is studied thoroughly. Based on the obstacle distance obtained from the vision algorithm, the brake and acceleration commands are sent from the on-board PC to the motor controller to avoid any collision with obstacle. The entire communication between the camera and buggy motor via the computer is established using Robotic Operating System (ROS) which is briefly introduced in the thesis. The limited space and highly dynamic obstacles were major challenges. The final software architecture was tested in real world and results were evaluated. The results show that the system is capable of detecting people irrespective of the pose, colour and illumination with small miss rates and false positives and control vehicle’s speed without considerable delay. The report will end with future suggestions.Master of Science (Computer Control and Automation