| This project initially completed a review of the different monocular techniques and summarized their suitability for obstacle detection on a moving vehicle. The Optical Flow method was chosen to explore further as the key technique for obstacle detection, based primarily on the work by Srinivasan (2000). A list of optical flow algorithms was created. The OFLib library was selected as the most appropriate algorithm for our application, primarily due to the fact that the library was designed to process on a GPU. A C;The application performance met expectations, as it was able to run in real-time' and created a mask of the obstacles in the original image. Further testing revealed that three major types of noise were present in the system. It was estimated that 30-50% percent of all outputs were significantly affected by at least one of the identified noises. Two of the sources of noise are caused by the turning and bouncing of the camera. The last type of noise was found in the ground plane estimate algorithm that was implemented. Each type of noise was analysed and a theoretical solution presented to minimize the effect. |
