Vehicle Ego-Motion Estimation by using Pulse-Coupled Neural Network

Abstract

This paper presents a visual odometer system using a monocular camera for vehicle navigation. A novel algorithm for vehicle ego-motion estimation based on optical flow and image segmentation is proposed. By applying a Pulse-Coupled Neural Network (PCNN), the image is dynamically divided into road area and non-road area by analysing texture smoothness. Correct road region detection effectively reduces computation cost and improves accuracy of egomotion estimation. Then a novel optical flow optimization method is proposed to produce reliable optical flow field in the road area detected previously. It's known when the vehicle is moving on a planar structured road, its 2-D motion field is expected to have specific form. Therefore ego-motion of vehicle, instantaneous speed and angular velocity, can be recovered from optical flow field of road area. Experiments show that the visual odometer successfully provides driver with robust and accurate vehicle self motion information.

Homography-based ground plane detection using a single on-board camera

“…However, this information may not be available in all applications, and even if available it might be inaccurate because of drifts [5,8], thus it is interesting and challenging to capitalise on all the information attainable via pure vision analysis. Unfortunately, visual ego-motion estimation is a complex task and still an active research line [8][9][10].…”

Vision-based vehicle detection and tracking with a mobile camera using a statistical framework

“…In this context, the use of inertial sensors (e.g., odometers, speedometers, accelerometers, gyroscopes) has been proposed in numerous works to facilitate ego-motion estimation (e.g., Stein et al 2000). However, this information may not be available in all applications, and even if available it might be inaccurate due to drifts (Simond, 2006;Cao et al, 2007), thus it is interesting and challenging to capitalize on all the information attainable via pure vision analysis. For instance, the authors in (Yamaguchi et al, 2008) estimate the ego-motion by computing the essential matrix between two consecutive views of a scene using feature correspondences.…”

“…In (Rabe et al, 2007) a Kalman filter based ego-motion compensation is proposed that merges stereo and optical flow information. Unfortunately, as stated, visual ego-motion estimation, especially in monocular systems, is a complex task and still an active research line (Cao et al, 2007;Gavrila and Munder, 2007;Ess et al, 2009).…”