Forward collision warning system using cascade classifier on embedded platform

Abstract

Every year, about 1.3 million people died on road traffic worldwide. Road traffic death number continues to increase steadily over years and reached 1.35 million in 2016. Road traffic injuries have become the leading killer or cause of death for children and young adults with range of age between 5-29 years old (WHO, 2018). Thus, traffic accident problem is crucial and need to be improved. Advanced Driver-Assistance Systems (ADAS) help reducing traffic accidents caused by distracted driving. One of the features of ADAS is Forward Collision Warning System (FCWS). In FCWS, car detection is a crucial step. This project explains about car detection system using cascade classifier running on embedded platform with single camera as the only sensor. Developing a system that use single camera to detect a car and predict its distance is quite a challenging task. Thus, a good image processing algorithm is needed to perform such task. The embedded platform used is NXP SBC-S32V234 evaluation board with 64-bit Quad ARM Cortex-A53. The system algorithm is developed in C++ programming language and used open source computer vision library, OpenCV. For car detection process, object detection by cascade classifier method is used. The cascade detector is trained using positive and negative instances mostly from our self-collected Malaysian road dataset. 4564 positive images have been used to train a car detector which turns out to be enough to produce acceptable results. A distance estimation method is used, and the developed system can give warning based on distance and Time to Crash (TTC). The distance estimation algorithm can produce acceptable results with maximum deviation percentage of 3.88875 %. In term of performance, when running on embedded platform, the FCWS runs at average of 9.52 FPS when scanning default ROI and at average of 16.86 FPS when tracking a car.