Building a low cost advanced driver assistance system : human detection

Abstract

Driving safety has always been a top priority in the automotive industry. Advance Drive Assistance Systems (ADAS) are systems which helps driver in their driving process can enhance manufacturers’ competitiveness in the market. In the past decade, more emphasis has been placed on using such systems on driverless vehicles. As such, the development of ADAS is highly sought after in the industry. ADAS comprises of multiple sub-systems to operate and this project will aim to look into the human detection system for ADAS.

For this project, the detection system making use of Histogram of Oriented Gradient (HOG) feature descriptor, alongside with Support Vector Machine (SVM) classifier would be studied. The detailed explanation on HOG and SVM would be covered in the subsequent chapters. HOG which are computed on a dense grid of uniformly spaced cells and overlapping local contrast normalisations assembles the features extracted into a histogram. SVM will thereafter make use of the histogram to classify the features and eventually computing bounding boxes around areas which are detected to have the object of interest (human). These detected bounding boxes would then be compared with ground truth bounding boxes to establish the amount of overlap that both types of boxes have. With the amount of overlap between each detection bounding box and ground truth bounding box calculated, the precision and recall value of the detections will then be computed. Then, the relationship between the overlap threshold, precision and recall would be established. This will then form the evaluation for the detection system.