Monocular-camera based Obstacle detection with Measurement Error Estimation

摘要

In this paper, we propose an obstacle detection method by using a monocular camera for a vehicle safety assistance system. The proposed method is based on 3D reconstruction given by sequential images and 3D information is used to detect obstacles and the distance to them. However, when the optical axis of the camera and the direction of vehicle movement are close, as in the case of a backward-looking camera, this causes a 3D reconstruction error which differs greatly for each location in the image. For this reason, detecting obstacles and finding their positions become difficult. The proposed method expresses the probability of obstacle existence by using an occupancy grid and taking into account the measurement error. Furthermore, we use sequential information to perform more stable detection. A complete system for obstacle detection from image sequences was constructed. Experiments show that the system is able to detect small targets in real-time.