Eureka! Institutional Repository

There has been significant progress in applying deep learning techniques to computer vision problems for perception scenarios, specifically for autonomous driving. This thesis explores some of these techniques and presents a detailed analysis of how deep learning tools can be used to perform computer vision for self-driving vehicles. It is an attempt to replicate concurrent research work, while discussing the advantages and disadvantages of different approaches. The thesis also combines different methods/components and presents a custom vehicle detection approach. The approach is based on generating a spatial grid of classifications, and then regressing bounding-boxes for pixels with a high object confidence score. The custom detection approach was tested on the KITTI object detection benchmark and was able to successfully detect objects of varying scale, lighting conditions and orientation. Additionally, an analysis of semantic segmentation techniques that use deep learning is presented and a few hand-picked approaches from literature are evaluated and compared. One of the approaches is replicated and tested on the Cityscapes benchmark for pixel level semantic segmentation. Detailed discussion and insights into future work are presented, which could be interesting for both academia and industry, especially in the area of deep learning and autonomous systems.