Fast Multi-line Detection and Tracking with CUDA for Vision-Based UAV Autopilot

This paper introduces an algorithm for fast multi line tracking utilizing the GPUs (Graphic Processing Units). Video stream contains huge of information for manipulating the vehicle by itself. It is a sort of big data to analyze properly in real-time for autonomous flight. However, image processing is heavy work for computing unit which is equipped on small unmanned aerial vehicle. This paper presents feasible image processing system for vision based intelligent vehicle. The proposed techniques for multi-line tracking are Hough transform, Kalman filter and clustering with GPUs. Integration of these methods has advantages to reduce the computational load by prediction of next state during the tracking and being robust for noise and rapid change of line's position. Hough transform used for extraction of lines while the Kalman filter predicts future state. Hough transform is easy to implement and robust for noise, on the other hand, the resource consumption raises exponentially as the resolution of input image or when we need high precision in Hough space. One of the efficient ways to overcome this speed problem is performing image processing with GPU's massive parallel calculation capabilities. Performance evaluations show promising results with acceptable trade-off between speed and accuracy of algorithm. Improving in speed algorithm keeps accurate tracking in comparison with algorithm implementation on CPU that is unable to track and detect lines fast enough due to computation resource limitations. Experiments and performance analysis of algorithm verified with user-made multi lines.