A Review on Advanced Detection Methods in Vehicle Traffic Scenes

In an age that is changing over time, vision-based technology is also constantly evolving. Object detection plays a vital role in these trends. This paper provides review information on vehicle detection methods used for vehicle traffic monitoring system. The detection methods are analyzed by extracting different features techniques which are mainly for multi-scale vehicle objects and various environment conditions. In rapid growth of deep learning methods for vehicle detection task, it is focused on region-based methods such as R-CNN (Region-based Convolutional Neural Network) and regression-based methods such as YOLO (You Only Look Once) and also presented each improved versions. In this survey, the performance analysis for detection methods are depicted with experimental results using benchmark datasets in measuring accuracy. This review paper is aimed at the improvement of advanced deep learning frameworks for vehicle detection in real-time.

I. Introduction

In modern technology age, object detection is a significant technique in computer vision field. Vision-based object detection task identifies and locates objects of a definite class, such as vehicles or pedestrians within an image or video. Detecting various objects in vehicle traffic scenes is become essential part for intelligent transportation system. Therefore, a lot of researchers is continuously innovated to improve the vehicle detection methods. But vehicle detection still facing the challenging problems because of vehicle appearance changes, multi-scale difficulties, occluded conditions, illumination variation and various environment situations. Researchers have addressed these issues through the use of deep learning-based methods. To perform the experiments, researchers are trained and tested on benchmark datasets such as KITTI, UA-DETRAC. The image data in these datasets are run by deep learning algorithms through a lot of layers of neural network algorithms, each of which passes a simplified representation of that data to the next layer.