The topic of the thesis is object detection in ADAS and autonomous vehicle systems. At first there is an overview of the sensors used in such vehicle systems, the emphasis is taken on radar, LiDAR, and camera based sensor systems. After the overview a sensor is chosen.
Later there is an examination of the most popular object detecting algorithms, both neural network and traditional feature based. The goal is to find an algorithm that is capable of real time perception, even on an embedded system.
After choosing the sensor and the algorithm to be used, the attention is drawn to different hardware platforms, with special attention to parallel architectures. After examining the CPU, GPU, and FPGA based platforms, a GPU based platform is chosen, and the work is described to make the algorithm run optimally on a GPU.
After porting and optimizing the algorithm to the new hardware platform, the achieved results are evaluated using both field tests and simulation measurements. It turns out that the GPU based implementation can run almost 10 times faster, achieving at least the same level of perception as the original heavily CPU optimized implementation.