부산대학교 도서관

Obstacle detection and drivable area recognition are time-critical operations in the motion planning of a self-driving vehicle. Computing Occupancy Grid Maps (OGM) from stereo image data is a promising solution because of its cost-effectiveness when compared to active sensor based solutions. The drawback however is that it is computationally intensive and takes more processing time. This work aims to accelerate the computation of OGM from stereo image streams by implementing parallel processing modules in the algorithm. This work also discusses the implementation of different disparity computation techniques in GPU and their impact on the performance of the overall algorithm. The objective is to improve the frame rate at which the stereo system is able to update the OGM while maintaining an acceptable level of accuracy. A frame rate of 4fps is achieved on a low-end Nvidia GPU for an image size of 1344x372, which gives 68% improvement over CPU-only solution.