Su-Birm Park

Vision-based rain sensing with an in-vehicle camera

An intelligent wiper speed adjustment system can be found in most middle and upper class cars. A core piece of this gadget is the rain sensor on the windshield. With the upcoming number of cars being equipped with an in-vehicle camera for vision-based applications the call for integrating all sensors in the area of the rearview mirror into one device rises to reduce the number of parts and variants. In this paper, functionality of standard rain sensors and different vision-based approaches are explained and a novel rain sensing concept based on an automotive in-vehicle camera for Driver Assistance Systems (DAS) is developed to enhance applicability. Hereby, the region at the bottom of the field of view (FOV) of the imager is used to detect raindrops, while the upper part of the image is still usable for other vision-based applications. A simple algorithm is set up to keep the additional processing time low and to quantitatively gather the rain intensity. Mechanisms to avoid false activations of the wipers are introduced. First experimental experiences based on real scenarios show promising results.

Vision-Based Pedestrian Detection -- Improvement and Verification of Feature Extraction Methods and SVM-Based Classification

Feature extraction and classification are two of the most important modules of any vision-based pedestrian detection system, since they are critical to the performance of the system as a whole. This paper presents the feature extraction and classification modules of a vision-based pedestrian detection system using a vehicle-mounted monochrome camera. The feature extraction module includes two kinds of features: wavelet-based features and a combination of simple symmetry and edge density features. Support vector machines based on a modified version of libSVM (Chang and Lin, 2001) are used for classification, and, for feature selection and optimization of feature space size, a fast and simple method using image masks for both feature types is presented. We have trained and tested our system using pedestrian and non-pedestrian images extracted from video sequences showing daylight urban traffic scenes

ProFusion2 — towards a Modular, Robust and Reliable Fusion Architecture for Automotive Environment Perception

This publication focuses on a modular architecture for sensor data fusion regarding to research work of common interest related to sensors and sensor data fusion. This architecture will be based on an extended environment model and representation, consisting of a set of common data structures for sensor, object and situation refinement data and algorithms as well as the corresponding models. The aim of such research is to contribute to a measurable enhancement of the output performance provided by multi-sensor systems in terms of actual availability, reliability, accuracy and precision of the perception results. In this connection, investigations towards fusion concepts and paradigms, such as ‘redundant’ and ‘complementary’, as well as ‘early’ and track-based sensor data fusion approaches, are conducted, in order to significantly enhance the overall performance of the perception system.

Motion segmentation using interest points

In this paper, we present a novel approach to motion segmentation by using interest points. Distinctive image features, so called interest points, were extracted in each image of the sequences and tracked using a Kalman filter. The interest points are then organized in a undirected graph. This connected structure can be used to describe the spatial relationship of interest points. An edge scoring algorithm is introduced that favors edges connecting interest points lying on the same object and punishing bridging edges, e.q. edges connecting different objects. To do so, we will introduce our so called ldquohomogeneous scale assumptionrdquo that is used to calculate scores for each edge in the graph. The resulting connected sub-structures in the graph are mathematically described by a radial map and then again tracked using a Kalman filter to further increase robustness. The presented algorithm is capable of working at 30 Hz and is thus feasible in a wide variety of applications.

A Symmetry Search and Filtering Algorithm for Vision Based Pedestrian Detection System

In this paper we present a fast symmetry search and filtering algorithm for monocular vision based pedestrian candidate detection application. First the ROI of symmetry search is focused on the pedestrian leg region, where the background is relatively simple ground plane. Afterward, the search region is divided into 2 x 4 sub blocks and symmetry density and distribution of each sub block is calculated. Finally, by comparing the symmetry density and distribution of the sub blocks, the correct symmetry axis of the pedestrian candidate is search and also some no-pedestrian candidates are filtered out. The results shown in this method are fast, cost effective and well suited for real-time vision applications.

A REAL-TIME DETECTION ALGORITHM FOR VISION-BASED PEDESTRIAN PROTECTION

This paper discusses the robust, real-time detection of stationary and moving pedestrians utilizing a single car-mounted monochrome camera. First, the system detects potential pedestrians above the ground plane by combining conventional Inverse Perspective Mapping (IPM)-based obstacle detection with the vertical 1D profile evaluation of the IPM detection result. Usage of the vertical profile increases the robustness of detection in low-contrast images as well as the detection of distant pedestrians significantly. A fast digital image stabilization algorithm is used to compensate for erroneous detections whenever the flat ground plane assumption is an inaccurate model of the road surface. Finally, a low-level pedestrian-oriented segmentation and fast symmetry search on the leg region of pedestrians is also presented. A novel approach termed Pedestrian Detection Strip (PDS) is used to improve the calculation time by a factor of six compared to conventional approaches.