VISION BASED ADVANCED DRIVER ASSISTANCE SYSTEM USING DEEP LEARNING

India is having one of the highest road casualty rates of almost nearly 1.5 lakh a year. The rate of road related deaths come down by a good margin if Advanced Driver Assistance System(ADAS) becomes mandatory for the vehicles. Since Indian car manufacturers focus on cost cutting, only Anti-lock Braking System(ABS) and airbags are offered as standard. With the advent of deep learning methods and computer vision, the use of an on-board camera and real time processing of captured images of road scenes makes it possible to achieve a wide range of ADAS features. In this paper a deep learning and computer vision based approach is used to develop a multi-functional ADAS system. The proposed system combines various objectives such as Vehicle Detection, Pedestrian Detection, Traffic Sign Board Detection, Traffic Light Detection and Blind-spot Vehicle Detection. The detection model is implemented using Tensorflow and various models are trained and tested using different detection algorithms and deep neural network(DNN) architectures. Networks such as SSD Inception, SSD Mobilenet, Faster RCNN Inception are trained on the same hardware set-up in order to perform a comparative study. The optimal model is selected based on a trade-off between detection time and accuracy. The model developed can detect 35 classes of objects that are commonly seen while driving a vehicle. The optimal model can be implemented on a suitable hardware and converted to a low-cost portable vehicle accessory.