Autonomous driving involves perceiving and interpreting a vehicle’s environment using various sensors for controlling the vehicle, marking drivable areas and locating pedestrians. A pedestrian detector plays a key role demanding real time response. An efficient pedestrian detector must determine the exact location of a pedestrian in complex backgrounds, poses, illuminations, due to which it is a source of an active research for the last two decades.
With the evolution of deep learning, there is no need of designing features which describe the pedestrian characteristics, instead the features can be learnt with the help of Convolutional Neural Networks (CNN). Our work includes training the YOLO v3 model on BDD100k dataset which is the largest and most diverse video dataset so far. It contains more pedestrian instances than previous specialized datasets, which makes it more viable for performing pedestrian detection. The results of training show that the proposed YOLO v3 network for pedestrian detection is well-suited for real-time applications due to its high detection rate and faster implementation.
Idea By: Aditya Sharma, Microsoft