Sam Schauland

A signal theoretic approach to measure the influence of image resolution for appearance-based vehicle detection

In this work a framework to measure the influence of training image resolution on classification performance for appearance-based object detection algorithms is presented. It is shown that based on sampling theory a reasonable image resolution for feature extraction can be chosen in advance, that is prior to the time consuming feature extraction and testing of the classifier. This is possible due to measuring the signal energy that is preserved in a low resolution image with respect to the optimal case of a high resolution image. The approach is justified using an AdaBoost algorithm with Haar-like features for vehicle detection. Tests of classifiers, trained with different resolutions, are performed and the results are presented. These results reveal that there is a good tradeoff between classification performance and computational load. The presented framework helps choosing a resolution for a good description of the training data.

Detection of Moving Objects in Image Sequences using 3D Velocity Filters

Detection of Moving Objects in Image Sequences using 3D Velocity Filters A movement analysis of objects contained in visual scenes can be performed by means of linear multidimensional filters, which have already been analyzed in the past. While the soundness of the results was convincing, interest in those systems declined due to the limited computational power of contemporary computers. Recent advances in design and implementation of integrated circuits and hardware architectures allow realizing velocity filters if the n-D system is carefully adapted to the analyzed problem. In this paper, the fundamental principles of visual scene analysis by linear multidimensional filters are examined with respect to possible sources of degradation. The extraction of movement information and its practical use are demonstrated using a wave digital filter (WDF) implementation.

Multisensor data fusion for advanced driver assistance systems - the Active Safety Car project

Driver assistance systems support overstrained and affected drivers and become more and more essential for series-production vehicles. Object detection and segmentation is one of the most challenging research topics in this field. In order to warn the driver or automatically break before a potential collision, objects intersecting the path of the host vehicle have to be detected and classified. Most recently developed approaches are based on two dimensional image processing, sometimes in combination with a tracking algorithm associating detections in consecutive frames to one and the same object. Further robustness is achieved by multisensor data fusion, i.e. information by two or more different sensors (e.g. camera and radar data) are fused in order to get a much more reliable result. Another aspect for safety applications is communication between cars, which provides additional sensor locations and thus also requires data fusion technology. Two different approaches for data fusion are proposed and first results are presented.

Motion-Based Object Detection for Automotive Applications using Multidimensional Wave Digital Filters

Vision-based object detection is a core part of many automotive collision warning systems. Especially due to the extensive information an image can hold and the decreasing costs of computational power, it has attracted more and more interest in the last decade. In this paper we propose a motion-based approach to simplify the detection of moving objects in order to improve available methods and make them more efficient. We interpret the image sequence containing the moving object (e.g. a vehicle or a crossing pedestrian) as a three-dimensional signal, not just as a sequence of image matrices. That way we can benefit from the advanced theoretical knowledge on system description and handling from the field of signal processing. In detail, three-dimensional velocity filters implemented using wave digital filters are used to oppress every object in the image that is not moving in a certain direction at a certain velocity.

Realization of a recursive 3-D cone filter for video processing applications

In this paper a realization of a recursive 3-D cone filter applicable for low cost hardware like FPGAs, for instance, is presented. Based on the filter design method proposed by Bolle, multidimensional wave digital filters are used to approximate a cone-shaped frequency response. Focus is not set on filter design but on filter implementation and problems arising in that context. In detail, the impact of signal size, chosen processing scheme and shift operators on hardware implementation is illuminated and a suitable solution is proposed. For verification, an example of a filtered video sequence showing moving objects is given.

3D wave digital filter implementation on a virtex2 FPGA board with external SDRAM

Many n-D signal processing applications require realization in real time. We propose the realization of a 3-D spatio-temporal wave digital filter (WDF) in an FPGA. Optimization of the implemented hardware architecture includes evaluation of two different kinds of overflow handling, namely by saturation and a “modulo 2” type operation. The FPGA board is processing DVI signals that can be provided by usual PC hardware. The processing output is observed in real time on an LCD monitor.

Motion-based object detection using 3D wave digital filters

This paper is focused on the separation of three-dimensional signals due to different velocity or directional components, more precisely detecting moving objects in visual scenes by means of linear multidimensional filters. While soundness of results obtained by similar systems developed in the past has been convincing, interest in those systems declined to the reduced computational power of contemporary computers. Recent advances in computation speed as well as design and implementation of integrated circuits and hardware architectures allow hardware realization of the above mentioned velocity filters for real-time processing. We present wave digital filter (WDF) structures adapted to the detection of objects moving in a certain direction at a given speed. These filters have been successfully tested on sequences showing crossing pedestrians recorded by a vehicle-mounted camera.

Multidimensional velocity filters for visual scene analysis in automotive driver assistance systems

Automotive scenery often contains objects that can be classified by object speed and movement direction. These features can be extracted from video data by linear n-D filters, which have already been analyzed in the past. While soundness of results was convincing, interest in those systems declined due to the reduced computational abilities of contemporary computers. Modern hardware allows realization of velocity filters, if the n-D system is carefully adapted to the analysis problem. The present paper analyzes the premises for application of velocity filters in the domain of automotive driver assistance systems, i.e. with respect to detectability of objects and implementability in a cost effective way. Especially the influence of the frame rate and the temporal violation of the sampling theorem are analyzed. Transfer functions for n-D filters working in a vision-based blind spot collision avoidance system are presented and discussed, and promising approaches for future application fields are proposed.

Tomographical scene reconstruction in the active safety car project

Tomographical reconstruction algorithms can be applied to camera based measurements and thus reconstruct the scenery without knowledge about included objects. The latter is interesting in the domain of driver assistance systems that have to monitor the driveway independently from a priori knowledge about possibly appearing objects. The paper presents tomographical background information, the transfer from radiographing to visual light rays and its negative impacts, and some first result obtained by applying the presented algorithm to images of a from view camera.

Realization of multidimensional signal processing systems given in natural k-D state space description

The paper addresses design, analysis, and realization of linear, discrete domain k-D signal processing systems based on a natural state space description. The latter is chosen to overcome restrictions by widely known Givone-Roesser and Fornasini-Marchesini models. It is capable of describing noncausal systems as well as shift operations with respect to more than one coordinate direction. It thus provides a system description that offers better compliance with possible hardware realizations than existing models and thus allows evaluation and consideration of many realization specific effects.