Basel Fardi

Shape and motion-based pedestrian detection in infrared images: a multi sensor approach

This work deals with the detection and tracking of pedestrians. The focus of the investigations was on methods, which allow a precise and detailed description of both significant features of pedestrians: shape and motion. Since the practical employment of such methods requires a good initialization and tracking, a multi sensor system was developed consisting of a far infrared camera, a laser scanning device and ego motion sensors. To handle the combination of the information of the different sensors a Kalman filter based data fusion is used. Arranging a set of Kalman filters in parallel, a multi sensor/multi target tracking system was created. The system structure combines a straightforward with a backward loop methodology to combine fast initiation functions with more affordable verification functions. Therefore formerly known semiautomatic image processing methods work fully automatically in the system. The analysis of the estimated optical flow regarding the typical human motion as well as the analysis of shape parameters using active contour models is performed. The multi sensor/multi target tracking system is installed on a test vehicle to obtain practical results, which are also discussed in this article.

A new approach for lane departure identification

This paper introduces a lane-departure identification (LDI) system of vehicles on a road with lane marks. The system perceives the LD by means of an artificial vision relying on an ad hoc approach like a boundary pixel extractor (BPE) to make up the robustness of machine vision. For the LDI,lane-related information composed of the direction and the position of lane boundaries in images is extracted. Then using the two parameters, it is determined whether or not a vehicle deviates from its lane. Except for the two parameters, the proposed system does not use any information such as lane width, a curvature,time to lane crossing, offset between the center of a lane and the center axis of a car body and other vehicle-related data. Besides, a camera calibration, any coordinate transformation, and a road model are not required. The system was demonstrated under various situations of changing illumination, in various road types without speed limits showing a quite good performance.

Obstacle Detection and Pedestrian Recognition Using A 3D PMD Camera

This paper presents a 3D-camera system and appropriate algorithms for the image processing to provide pedestrian recognition. According to international legislative proposals the automotive industry is forced to take action in the area of protecting vulnerable road users. The presented photonic mixer device (PMD) sensor system is able to fulfil specific requirements of a pedestrian protection assistant. It consists of a sensor with a resolution of 64 times 16 pixel, lightsources, and the image processing unit. The light emitting diodes (LEDs) of the lightsources emit a modulated signal in the infrared (IR) spectrum. The sensor calculates the object distance by means of the phase of the reflected signal. The range of the system is approximately 15 m for pedestrians and up to 30 m for objects with high reflectivity, e.g. cars with number plates. The horizontal field-of-view is 55deg. The presented image processing unit consists of two main steps. In the first step both a robust and efficient segmentation method is performed to create reliable detections of the objects and a useful description of their projection in the image plane. It uses the linking pyramid method regarding the partition of the distance image and a contour-based grouping algorithm, where the objects are described using the chain code representation. In the second step a classification of the resulting objects is carried out. Depending on the distance of the objects which have to be classified a shape- or motion-based verification is applied adaptively. Both approaches discussed in this paper deliver very good results at the corresponding distance and represent a solid foundation of further works

Multi-modal detection and parameter-based tracking of road borders with a laser scanner

This paper treats an important problem concerning driver assistance systems: the detection and tracking of road borders. The detection of the road boarders are created from the signals of a laser scanner system. Two different information can be taken from the laser signal: range and reflectivity. The range signal delivers the road edges at the basis of single scans. In opposite to that, to detect edges in the reflectivity, the reflectivity signals are arranged as images and a special image processing algorithm is developed. The fusion and tracking of this information is performed using an Extended Kalman filter where a circular curve is used as movement model.

Hough transformation based approach for road border detection in infrared images

This paper describes a real-time image processing algorithm for road border detection in infrared images. The basic idea is to couple the road border lines through the parallelism between them in the vehicle coordinate system. The parallelism of the two lines is translated in their convergence in the image plane due to the perspective projection. In this plane the two lines meet at a point on the horizon line. The new idea is to use this common point on the horizon line to find the road border in the Hough domain. The input data for the Hough transformation is created by a set of local regularized edge detectors and an adaptive thresholding of the image. Since the image mostly shows low contrast, the edge extraction can hardly be realized by edge detectors with a small kernel. The image, therefore, is subsampled using the Gaussian pyramid technique and the preprocessing takes place in an optimal resolution level that is experimentally determined. The developed algorithm has been implemented on an experimental vehicle equipped with an infrared camera and was successfully tested in different situations.

Multi sensor based tracking of pedestrians: a survey of suitable movement models

This article presents a multi sensor approach for driver assistance systems: the detection and tracking of pedestrians in a road environment. A multi sensor system consisting of a far infrared camera and a laser scanning device is used for the detection and precise localization of pedestrians. Kalman filter based data fusion handles the combination of the sensor information of the infrared camera and of the laser scanner. Arranging a set of Kalman filters in parallel, a multi sensor/multi target tracking system was created. The usage of suitable movement models has a great influence on the performance of the tracking system. Several types of models are discussed focussing on the typical behavior of pedestrians in road environments. The multi sensor/multi target tracking system is installed on a test vehicle to obtain practical results which is discussed in this article too.

Free space determination for parking slots using a 3D PMD sensor

We present an approach for parking slot detection using a 3D Range camera of PMD type. This sensor allows referring to a large number of spatial point measurements detailed representing cuts of the observed scene. The focus of this paper is on the feature extraction out of the PMD data as well as the fusion of the features defining the free space of a parking slot. The feature extraction includes reliable and exact curb detection and robust obstacle detection. The approach for the optimal feature extraction is based on the usage of an occupancy grid in combination with a feature conform definition of detection channels.

Motion-based pedesvtrian recognition from a moving vehicle

This article presents the results of a detailed investigation of motion-based recognition of pedestrians. Here, the influence of the ego motion and its compensation in view of the recognition performance play a central role. Therefore, two basic approaches for the estimation of the object movement are specified, tested and evaluated. Moreover, different features for the generation of the time series are analyzed. These form the basis for the recognition of the periodicity of the human walk. The time series are used for the generation of a power density spectrum, in which the verification of the object movement is carried out. Since the quality of the evaluated spectrum and hence the robustness of the recognition depends on the duration of the measurement sequence, this correlation is carefully considered in order to define the marginal conditions. Furthermore, all variations of different walking directions are tested, as well as the distances of the object in relation to the position of the own vehicle. The obtained results show that in most cases the recognition of a moving human being is possible. This method represents an interesting possibility to increase the reliability of technical systems for recognizing of pedestrians

Road Border Recognition Using FIR Images and LIDAR Signal Processing

This paper addresses the detection and tracking of road borders in non cooperative environments. A 2-dimensional scanning LIDAR is used to improve the reliability of the FIR camera based road border recognition. In order to detect the road boarders we apply a Kalman-filter based model fitting strategy. Extracted measurements of the FIR images are transformed into the vehicle coordinate system in order to provide a precise description of the road course ahead. The model description in a common coordinate system -the vehicle coordinate system -allows an easy compensation of the ego-motion and a direct and straight forward fusion of the different sensor data. Both the detection and the estimation were developed and enhanced for the intended sensor configuration. The corresponding mathematical derivations are presented in this paper. Using the range values, delivered by the LIDAR, a more stable estimation of the pitch angle can be achieved. The realization of that is shown in detail. This is used to define the ROI in which the image processing is carried out.

Non-rigid-motion recognition using a moving mono camera

The focus of this contribution is the detection of moving objects and the classification of their motion as rigid or non-rigid. A new processing approach for analyzing video data of a moving monocular camera is introduced. It is mainly based on the evaluation of the optical flow in two different processing steps: Detection and classification. First, the ego motion is evaluated in order to separate moving objects from the background clearly. The implemented corresponding algorithm features both, the extraction of wrongly estimated displacement vectors and the ego motion compensation. The processing step applied after that deals with the classification of the detected objects. It will be shown that a reliable distinction between rigid and non-rigid objects is well realizable using some features derived from the motion orientation histogram. For a corresponding evaluation, the case of the pedestrian recognition was taken into account.