Depth estimation using single fish-eye camera

Abstract

Nowadays, with the increasing application of automotive navigation systems on automotive platforms. The robot system is required to measure the distance of the surrounding area in real-time. The omnidirectional distance measurement technology is in significant demand. Traditional distance measurement methods for example stereo cameras and time of flight sensors have a limited field of view. A scene with 360° field of view (Hereafter FOV) often requires multiple sensors which not only increases the cost but also increases the processing time.

Fisheye lens perception can be an effective way to use for environmental information collection. The FOV of a single fisheye camera can provide a hemisphere view of 180°. Theoretically, 2 fisheye camera modules have the capability to cover 360° FOV, avoid visual blindness and greatly improve the processing time with two fisheye camera modules.

Traditional image distance estimation methods are used to estimate the distance by objects’ given size and the camera’s focal length. A scene with multiple types of objects is required a large database to store the object image and size, which increases the system cost and processing time.

This research is about identifying a fast and robust distance measurement method by using a single fisheye camera. The technique can be able to cover the hemisphere view of 360°. This method calculates the target distance by using the ellipse geometry model of the fisheye camera.

The method first establishes an ellipse geometry model and obtains the fisheye camera parameter with a simple and easy-to-use calibration method. The object detection algorithm is applied to recognize the target object in the image, and the object points are fed into the model. Finally the target position, distance and size can be calculated by the equation.

In this research, the distance estimation method is developed. This method does not require inputting any dimension of the target before the measurement. This method is developed as a human safety detector of the robot. The human distance is estimated in the image. When the human is located within the danger zone, the system stops the robot.

In the experiment, 2 humans with different heights are selected. They are sitting and standing in front of the system. Over 900 sets of data are collected. The experimental results show that this method can effectively identify the target in the scene and extract their position and size. The measurement precision is greater than 90%. The accuracy does not affect by the target size and pose. The improved distance measurement method can be used for robotic applications, for example, real-time obstacle avoidance.