Towards solving perception based autonomous driving assistant system

Abstract

This thesis scrutinizes the problem of perception in the self-driving car system. Selfdriving car is the face of the future and the decade’s research focus. Tech giants like Google, Uber, Tesla, Commai, Intel MobilEye etc. are now immensely investing in this particular technology. In our work, we mainly address the perception problem of autonomous vehicle and try to solve it with only cameras and comparatively lower computational cost. Firstly, to detect the lane we propose QLD (Quick Lane Detection) model on CULane dataset which gives significantly improved results in the roads of countries like Bangladesh than other existing methods. Secondly, for object detection we propose our own dataset BDCO or Bangladeshi Common Objects, and merge it with MS COCO dataset to make it suitable for Bangladeshi roads. We train BDCO dataset in a CNN based object detection model (CbOD) which also gives very promising results in local roads. Finally, we cascade QLD and CbOD with our decision-making system which outputs the warnings based on the analysis of the inputs from cameras in the vehicle. Our hands-on evaluations show that, our cascaded network Bangladeshi Driving Assistant (BD-DA) attains performance competitive to the state-of-the-art systems on a indistinguishable benchmark.