A Study on Semantic Segmentation for Autonomous Vehicles

Autonomous vehicles are already a reality, and there are still several challenges to overcome. One important challenge for the adoption of these vehicles is perceiving its surroundings. This necessity of perception can be fulfilled by digital cameras. When working with digital image processing, the quality will be limited by real-time constraints. As several works indicate, this real-time constraint for autonomous vehicles is at most 100ms per frame. Also, by improving the processing time, the chances of accidents involving autonomous vehicles may be decreased. This paper analyses the advantages and drawbacks of semantic segmentation and also presents a study to implement perception for autonomous vehicles by accelerating a semantic segmentation algorithm, also used by other works on the field. To accelerate the algorithm, spacial parallelism will be used.