Real-time moving object detection using a vehicle-mounted monocular rear-view fisheye camera

Abstract

In this paper, we propose a moving object detection algorithm that enables to detect moving objects with collision risk in rear of the vehicle by using image sequences from a vehicle-mounted monocular rear-view fisheye camera.

The proposed moving object detection algorithm detects corner points by using Harris corner detector and computes the optical flow vectors from two consecutive images corresponds to the detected corner points. By considering the feature of the vehicle movement that the vehicle goes straight in a short time interval, we find the focus of expansion (FOE) by using matched filter and divide the image into four sections around the FOE. The optical flow angle distribution of each section is analyzed to find pixels corresponds to the background components and robust background motion compensation method to the complex scene is suggested.

Add to this, we propose false removal method that eliminates false positives by considering two features of the detection box for the moving object candidates; position and pixel intensity distribution. For more accurate object detection, we use the mean shift tracking algorithm simultaneously with detection algorithm.

Simulation results show that our proposed algorithm achieves 97.19\% of detection rate toward various detection target including pedestrians, bicycles, and cars. Furthermore, our proposed false removal algorithm performs an extremely low 2.7\% of false rate toward false positives such as trees, shadows, and road markers.

본 논문은 움직이는 단안 카메라 환경하에 물체를 탐지하기 위한 알고리즘을 제안한다. 단안 카메라에서 출력되는 2차원 영상을 Feature point, Optical flow, Focus of expansion의 기법을 사용하여 배경과 물체를 구분한다. 탐지한 물체를 바탕으로 Mean shift tracking을 병행하여 물체를 찾고 추적한다.