DEEP LEARNING-BASED VEHICLES AND PEDESTRIAN DETECTION IN INTELLIGENT DRIVER ASSISTANCE SYSTEMS

Abstract

In order to build a traffic safety system, we first need to develop ADASs. These systems normally
require real-time and reliable detection performance. However, moving vehicles and pedestrian detection
is critical requirement due to their challenges in the real-world environments such as complicated
background, shadow, partial occlusion, articulation and illumination variations. Besides, one of the most
important challenges in ADASs is real-time requirement. This paper proposes a model using deep-learning
algorithm in order to increase accuracy and processing time for ADASs. Accordingly, we first propose the
YOLO (You Only Look One) model, moreover in order to improve detection performance we add sample
datasets for traing model. Experimental results are then conducted in a NVIDIA Jetson TX2 embedded
computer. Achievable results prove that the proposed work can speed up processing time of at least 1.6x
with detection rate of 90% for static cameras; and speed up processing time of at least 1.36x with detection
rate of 90% in high resolution images (1280x720 pixel) for moving cameras.