Sébastien Ambellouis

Obstacles Detection on a Road by Dense Stereovision with 1D Correlation Windows and Fuzzy Filtering

In this paper, we propose an original approach to obstacles detection based on stereovision with mono-dimensional correlation windows. The result of the algorithm is a dense disparity map associated with a confidence map. For each pixel, correlation indices are computed for several widths of windows and several positions of the window centre. Three criteria, extracted from each correlation curve, are combined by a fuzzy filter to define a confidence measure. Our 1D method is compared to a classical 2D method and shows better results in term of errors and density rate. In the context of obstacle detection, we show that a basic segmentation of our disparity map yields a better detection and marking of the obstacles. The method is validated on synthetic image sequences and our results are compared with those obtained using a classical 2D method

Fisheye-Based Method for GPS Localization Improvement in Unknown Semi-Obstructed Areas

A precise GNSS (Global Navigation Satellite System) localization is vital for autonomous road vehicles, especially in cluttered or urban environments where satellites are occluded, preventing accurate positioning. We propose to fuse GPS (Global Positioning System) data with fisheye stereovision to face this problem independently to additional data, possibly outdated, unavailable, and needing correlation with reality. Our stereoscope is sky-facing with 360° × 180° fisheye cameras to observe surrounding obstacles. We propose a 3D modelling and plane extraction through following steps: stereoscope self-calibration for decalibration robustness, stereo matching considering neighbours epipolar curves to compute 3D, and robust plane fitting based on generated cartography and Hough transform. We use these 3D data with GPS raw data to estimate NLOS (Non Line Of Sight) reflected signals pseudorange delay. We exploit extracted planes to build a visibility mask for NLOS detection. A simplified 3D canyon model allows to compute reflections pseudorange delays. In the end, GPS positioning is computed considering corrected pseudoranges. With experimentations on real fixed scenes, we show generated 3D models reaching metric accuracy and improvement of horizontal GPS positioning accuracy by more than 50%. The proposed procedure is effective, and the proposed NLOS detection outperforms CN0-based methods (Carrier-to-receiver Noise density).

A Colour Correlation-Based Stereo Matching Using 1D Windows

In this paper, we propose an original approach to colour correlation-based stereo matching with mono-dimensional windows. The result of the algorithm is a quasi-dense disparity map associated with its confidence map. For each pixel, correlation indices are computed for several widths of windows and several positions of the current pixel. Three criteria, extracted from each correlation curve, are combined by a fuzzy filter to define a confidence measure. A basic decision rule computes the disparity value and its associated confidence for most of the image pixels. A first study shows results obtained on grey level images with our 1D method and a classical 2D method. The method is applied to the RGB colour space: three disparity maps are computed and fused to compute the final disparity map. The method is validated on the Tsukuba image pair. On the first hand, we show that our method presents lower error rates with the RGB colour space than with the grey level image for identical density rates. On the other hand, our results are compared with those obtained using similar colour 2D methods (presented on the Middlebury Website). Our algorithm is ranked in the first places for each area of the image.

Transport system architecture for on board wireless secured A/V surveillance and sensing

This paper describes the system architecture set up by the consortium of the EUREKA CELTIC BOSS project for enhancing the security of passengers inside commuter trains. The functional approach, together with obtained technical improvements in the three domains of wireless communications, abnormal events detection and video compression and robustness enhancement are presented. The demonstrator set up in the project, which was installed in a real commuter train in commercial operation, is also reported as proof-of-concept.

Velocity selective spatio-temporal filters for motion analysis in image sequence

Toward real-time implementation of optical flow methods, we present an IIR filter structure suitable for local motion analysis. The motion analysis framework is divided into three steps. A set of velocity tuned filters is first applied to the image sequence. Then, for each pixel, the filter which gives the maximal response is selected. Finally, a segmentation step allows a clustering of pixels according to the previously computed motion index. We only focus on the first step of the method. We present the structure of the proposed IIR filters and show the influence of the parameters on their selectivity.

Toward a Real Time View-invariant 3D Action Recognition

In this paper we propose a novel human action recognition method, robust to viewpoint variation, which combines skeleton-and depth-based action recognition approaches. For this matter, we first build several base classifiers, to independently predict the action performed by a subject. Then, two efficient combination strategies , that take into account skeleton accuracy and human body orientation, are proposed. The first is based on fuzzy switcher where the second uses a combination between fuzzy switcher and aggregation. Moreover, we introduce a new algorithm for the estimation of human body orientation. To perform the test we have created a new Multiview 3D Action public dataset with three viewpoint angles (30°,0°,-30°). The experimental results show that an efficient combination strategy of base classifiers improves the accuracy and the computational efficiency for human action recognition.

Obstacle Detection Using a Single Camera Stereo Sensor

In this paper we present a catadioptric stereovision system based on a single camera, two plane mirrors and a prism mounted on a rotating axis. On the one hand, the system is able to project a real scene on two half images. Thus, contrary to classical stereo system, it deals with the problem of the camera synchronisation. On the other hand, the optical axis of the system can be steered to point a region of interest of the real scene. It has been developed in the context of on-road vehicle applications and more specifically for long-range road surveillance. This degree of liberty is useful when the region of interest is far away in the front of the vehicle because it requires a long focal length to reach a sufficient resolution. This optical system is associated to a dense stereo matching algorithm and a disparity map segmentation process. We present some results we have obtained using synthetic stereo images to illustrate the functionality of this setup.

Launch these Manhunts! Shaping the Synergy Maps for Multi-Camera Detection

We present a method for multi-camera people detection based on the multi-view geometry. We propose to create a synergy map by the projection of foreground masks across all camera views on the ground plane and the planes parallel to the ground. This leads to significant values on locations where people are present, and also to a particular shape around these values. Moreover, a well-known ghost phenomena appears i.e. when these shapes corresponding to different persons are fused then the false detections are also generated. In this article, the first improvement is the robust detection of the candidate detection locations, namely keypoints, from the synergy map based on a watershed transform. Then, in order to reduce the false positives, mainly due to the ghost phenomena, we check if the particular shape, for an ideal person, is present or not. This shape, that is different for each location of the synergy map, is generated for each keypoint, assuming the presence of a person, and with the knowledge of the scene geometry. Finally, the real shape and the synthetic one are compared using a similarity measure that is similar to correlation. Another improvement proposed in this article is the use of unsupervised clustering, performed on the measures obtained at all the keypoints. It allows to automatically find the optimal threshold on the measure, and thus to decide about people detection. We have compared our method to the recent state-of-the-art techniques on a publicly available dataset and have shown that it reduces the detection errors.

SINGLE-CAMERA STEREOVISION SETUP

The stereovision sensor described in this article has been developed during a research project called RaViOLi, for “Radar and Vision Orientable, Lidar”. The main outcome of this project is the improvement of driving safety thanks to the analysis of redundant data coming from several cooperative sensors installed on an autonomous vehicle. One is a high precision stereovision sensor whose field of view can be oriented toward a region of interest of the 3D scene, like the road in front of the vehicle at long distance. The sensor is composed of a single camera, of two lateral mirrors, and of a prism rotating about its edge. The mirrors project the left and right images of the stereo pair onto both halves of the imaging surface of the camera, yielding the equivalent of two virtual cameras with parallel axes. Rotating the prism changes the orientation of both optical axes while keeping them parallel.

Exploiting 3D geometric primitives for multicamera pedestrian detection

In this paper, we present an approach for multicamera pedestrian detection exploiting the concepts of multiview geometry and the shapes of 3D geometric primitives. Multicamera occupancy maps provide peak responses corresponding to the object detection but suffer from several false detections known as ghosts. The novelty of this paper is the introduction of shape patterns which can model the objects, such as pedestrians, by defining a kernel function in the projected occupancy space. This kernel depends upon the geometry of the 3D primitives and also varies in relation to their position with respect to the cameras in the real world configuration. For multiple objects visible across several cameras, we define a formation model which is the convolution of this spatially varying kernel with the set of possible object locations. The locations corresponding to detections can thus be obtained through a deconvolution process. For efficient computations, we further propose an estimated deconvolution process specific to our kernel responses which can also be heavily parallelized. We show the application of this process towards pedestrian detection by studying various 3D cylindrical primitives. Experiments on two public dataset sequences, including comparison with another approach, show the efficiency of the proposed method in terms of pedestrian detection and ghost pruning, including in adverse and challenging conditions.