Learning Approach with Random Forests on Vehicle Detection

A Close-up Monitoring System (CMS) has been designed in our research laboratory, which aims at avoiding any potential collision risk by detecting the frontal train's distance from the captured video. Histogram of orientated gradient (HOG) has been used as a feature descriptor, because it gives robust performance in various illumination conditions. Random forest algorithm is a conventional machine learning tool, but it is new in the driving assistant application. Besides, the predicting process of our classifier is very fast because it only depends on a limited number of simple tests in each randomly-trained decision tree. Based on the HOG features and random forest algorithm, a close-range train detector has been designed. This proposed detector works as one detection module in CMS, and the correct detection rate of the close-range train was nearly 100%, which means there was no miss-detection in our control experiment. Compared with the traditional non-learning method, our learning-based approach achieves much stronger recognition ability with less false alarms.

I. Introduction

Over the past few years, the driving assistant system has attracted great attention, because it can provide an additional guarantee for operational safety. In the event of collision in rail transportation, the casualties and economic losses are usually very heavy [1]. Twenty-five people were injured in a single collision accident in Singapore at the end of 2017 [2]. Over forty-nine persons were killed and a great number of wounded in a train collision in Egypt in August 2017 [3]. Therefore, the rail transportation generally needs a higher degree of traffic safety than road transportation.