Improved pedestrian detection and distance detection with stereo camera using computer vision

Abstract

Advanced Driver Assistant System (ADAS) is an advanced technology that provides assistance in driving and parking functions. The main purpose of ADAS is to increase road safety and enhance comfort of the driver. ADAS system generally is implemented along with sensors and cameras, to detect obstacles or pedestrian and respond accordingly. ADAS had been proven to reduce road fatalities by minimizing human error. One of the advanced systems used for ADAS nowadays is stereo vision cameras, which is set up using two identical cameras to produce a 3D image. The main highlight of a stereo vision is that a 3D image has depth which gives higher accuracy compared to a 2D. Nowadays, cases of road accidents have been increasing every year. According to World Health Organization (WHO), approximately 1.35 million people die each year due to road accidents. Among these number of cases, more than half of the deaths were among pedestrians, cyclists, and motorcyclists. Meanwhile currently most of the detection was done on sunny day, as it is much more challenging to detect pedestrian during complicated environment such as rainy day. The flow of the project is firstly to set up a pair of stereo cameras on the dashboard of a stationary car to detect pedestrians. Using OpenCV with Python language, stereo matching is completed to obtain a disparity map and depth map. Once the depth map is obtained, distance can be calculated. To increase the accuracy on detecting pedestrians, Haar Cascade classification was implemented. After that, a major image training, testing and validation is performed. In this thesis, a detection of pedestrian in rainy day environment is implemented using stereo vision cameras. This work had proved that depth extraction using the custom stereo camera can work in a rainy day with an error of 0.2m accuracy off and Haar Cascade classification was successfully implemented to detect pedestrians.