Vehicle Detection in Deep Learning

Teoh, Per Nian (2019) Vehicle Detection in Deep Learning. Final Year Project, UTAR.

Preview

PDF Download (4Mb) | Preview

Abstract

Robust and efficient vehicle detection is an important feature to utilize in the smart transportation system. With the development of computer vision techniques and accessibility of large-scale traffic transport data, deep learning has been enabled to on-road vehicle detection algorithms. In addition, traffic transportation system involves death and life concern which requiring high accuracy to ensure safety, also, the detection system for autonomous driving requires real-time inference speed in order to guarantee prompt vehicle control. In this report, a brief concept of training a deep CNN and how deep CNN works in object classification and localization is presented. The objective of this project is vehicle detection with deep learning, so, vehicles data set from highway, urban road and housing area had been collected and applied to the deep learning and computer vision algorithms. Due to the limited resources for training large-scale data set, the detecting classes will be limited to car, bicycle and motorcycle. Each class has roughly same amount of training images with each other. Some experiments have been conducted in this project to figure out which batch size performing well in the training process. Moreover, output of the convolutional layers has been visualizing for better understanding in CNN working principal. Finally, the result the vehicle detection performance in this project still have room from further improving, and a higher accuracy performance can be easily achieved by acquiring adequate data set and find the suited hyper-parameters to train the model.

Actions (login required)