Catalin Golban

Fast vision based ego-motion estimation from stereo sequences — A GPU approach

Visual odometry has been an important research activity in the last three years. Because the results of ego-motion estimation tasks are used in complex systems which need to work real-time, the motion estimation itself need to perform faster than real-time such that the remaining time slots can be used by other algorithms running on the same hardware. The main contribution of this paper is the implementation of a GPU based method for 3D ego-motion estimation. We identified the visual odometry method that is the best candidate for parallelization and we describe the details of the parallel implementation. We also present different tests performed on various traffic scenes to show the robustness of the method and the performance compared to the sequential implementation.

Vision based three-dimensional vehicle motion detection by minimizing nonlinear functions

This paper presents a method to accurately determine the three-dimensional motion of a vehicle based on image pairs acquired with a stereo system. The basic idea is to express the relationship between correspondent features that belong to non moving objects in successive frames as a nonlinear function of rotation and translation parameters. Minimizing the distance between the previously mentioned parametric correspondences and the correspondences computed using optical flow methods, leads us to accurate estimation of the motion parameters.

Improving accuracy for Ego vehicle motion estimation using epipolar geometry

This paper presents an original method for increasing the accuracy of ego vehicle motion estimation using video data. Our algorithm takes as input a monocular video sequence on which originally combines procedures for feature detection and filtering, optical flow, epipolar geometry and estimation of the rotation from the obtained essential matrix. Imposing a movement constraint on the rotation matrix, we obtain a powerful method for estimating the rotation of the vehicle from frame to frame. Furthermore, the obtained rotation and stereo data are used for computing the translation of the vehicle. The use of stereo data only for translation estimation diminishes the influence of stereo errors on rotation matrix. Experiments have been performed using various urban traffic scenes, with horizontal and vertical curvatures revealing a high degree of accuracy compared to reference measurements.

A practical method for ego vehicle motion estimation from video

This paper presents an original and practical method for estimating the rotation and the translation of a vehicle using video data. Imposing a movement constraint on the rotation matrix, we obtain a powerful method for estimating the rotation of the vehicle from frame to frame. Our algorithm takes as input a monocular video sequence on which originally combines procedures for feature detection and filtering, optical flow, epipolar geometry and estimation of the rotation from the obtained essential matrix. Furthermore, the obtained rotation and stereo data are used for computing the translation of the vehicle. Experiments have been performed using various urban traffic scenes which contain both curves and straight line roads.

Linear vs. non linear minimization in stereo visual odometry

Visual odometry has been an important research activity in the last two years and it has lead to numerous papers being published. Few surveys and comparative studies between approaches exist at the moment. This paper makes a comparative study between two visual odometry methods from accuracy, speed, sensitiveness to noise, and degree of parallelization points of view. The comparison is performed strictly from the perspective of minimizing the cost function, since this is one of the most critical steps in motion estimation from visual data. We also proposed a method based on Kalman filtering to achieve better accuracy in the presence of illumination changes based on correlating the measurement model noise with the intensity variations in time.

Speed estimation for scene objects using stereo visual odometry methods

This paper proposes a novel method to determine the speed of the surrounding vehicles in traffic scenarios. Relying on the video information obtained from a stereo camera mounted on a moving vehicle, we first determine the vehicle ego motion based on static scene features then we determine the relative motion between objects based on features situated on the moving objects. For robustness to false feature matches everything is plugged into a multi-RANSAC framework. The novelty of the method consist in the fact that the relative motion between the objects can be determined with the same algorithm that was previously used for ego motion estimation, the only difference consisting in the geometric constraints that are imposed to the subset of point features considered for inliers set detection and evaluation. Also, the proposed method does not rely on the fact that objects are detected previously and it does not detect the objects.

Direct formulas for stereo-based visual odometry error modeling

Visual odometry is the most suitable method for recovering the camera motion in the context of video processing applications. The main advantages it brings are the accuracy of the estimation, the computation efficiency, and the elimination of the need to synchronize a video processing system with other odometry sensors. There is a large amount of recently published visual odometry methods, but none of them provides a reliable error model for the estimation. The goal of this paper is to present an analytical method to compute the covariance matrix for a stereo-based visual odometry method and to analyze its performance and quality by using both synthetic data and real world video sequences acquired in urban traffic scenarios.

Moving rigid objects segmentation in 3D dynamic traffic scenes using a stereovision system

This paper proposes a novel method for detecting the moving vehicles in dynamic urban traffic scenes using a stereo camera. Relying on the fact that a set of feature points on a rigid 3D scene object are staying in a rigid 3D configuration, we propose to compute the relative motion between the camera and a moving object with an algorithm that follows from the visual odometry based motion estimation methods. Subtracting the camera motion we obtain the absolute object motion. Additionally we create a compact representation of the scene using superpixels computed from intensity and depth information. A graph-like structure is built, having superpixels as nodes and indicating neighboring relationships between adjacent superpixels. Objects are segmented using a fast region growing algorithm that considers as seeds the features used to compute the object motion.

On-Board 6D Visual Sensor for Intersection Driving Assistance

The problem of on-board intersection perception and modeling is complex and requires a wide field of view, dense and accurate data acquisition and processing, robust object detection and classification and fast response time. This goal can be achieved by using a large and redundant set of heterogeneous sensors and by fusing their information. Among the on-board sensors the visual sensors have the following main advantages: they are passive, and they provide the highest volume of information. The use of a pair of visual sensors in stereo configuration opens not only the possibility to infer the 3D coordinates for any image point but also the possibility to compute the 3D motion vector for any pixel. The exploitation of the motion information in driving assistance systems requires the estimation of the ego motion. This paper presents the architecture, implementation and use of a powerful on-board 6D visual sensor for intersection driving assistance.

An experiment on relative rotation estimation from distant points with monocular vision

This paper proposes a method to determine the relative rotation between two images acquired with a camera. It is considered that the camera is calibrated, and that the relative motion between the two images is small. The method is appropriate for a camera mounted on a moving vehicle and it is proven that the changes in yaw, pitch and roll angles can be accurately determined in this setup. We propose a RANSAC process that selects the distant points based on a new motion model for image pixels valid only for the points at infinite. Additionally, robustness of the method is increased by considering the fact that image deformations are 0 for the points at infinite.