Stefan Gehrig and Alexander Barth and Nicolai Schneider and Jan Siegemund, "A Multi-Cue Approach for Stereo-Based Object Confidence Estimation", In Intelligent Robots and Systems (IROS). Vilamoura, Portugal, pp. 3055 - 3060. 2012.

• Abstract
• BibTeX
• DOI

In this contribution we present an approach to compute object confidences for stereo-vision-based object tracking schemes. Meaningful object confidences help to reduce false alarm rates of safety systems and improve the downstream system performance for modules such as sensor fusion and situation analysis. Several cues from stereo vision and from the tracking process are fused in a Bayesian manner. An evaluation on a 38,000 frames urban drive shows the effectiveness of the approach compared to the same object tracking scheme with simple heuristics for the object confidence. Within the evaluation, also the relevance of occurring phantoms is considered by computing the collision risk. The proposed confidence measures reduce the number of predicted imminent collisions from 86 to 0 maintaining almost the same system availability.

@inproceedings{Gehrig2012Multi,
  author = {Gehrig, Stefan and Barth, Alexander and Schneider, Nicolai and Siegemund, Jan},
  title = {A Multi-Cue Approach for Stereo-Based Object Confidence Estimation},
  booktitle = {Intelligent Robots and Systems (IROS)},
  year = {2012},
  pages = {3055 -- 3060},
  doi = {10.1109/IROS.2012.6385455}
}

Alexander Barth and Jan Siegemund and Annemarie Meißner and Uwe Franke and Wolfgang Förstner, "Probabilistic Multi-Class Scene Flow Segmentation for Traffic Scenes", In Pattern Recognition (Symposium of DAGM). Goesele, M. and Roth, S. and Kuijper, A. and Schiele, B. and Schindler, K. (Eds.), pp. 503-512. Springer. 2010.

• Abstract
• BibTeX
• DOI
• Download

A multi-class traffic scene segmentation approach based on scene flow data is presented. Opposed to many other approaches using color or texture features, our approach is purely based on dense depth and 3D motion information. Using prior knowledge on tracked objects in the scene and the pixel-wise uncertainties of the scene flow data, each pixel is assigned to either a particular moving object class (tracked/unknown object), the ground surface, or static background. The global topological order of classes, such as objects are above ground, is locally integrated into a conditional random field by an ordering constraint. The proposed method yields very accurate segmentation results on challenging real world scenes, which we made publicly available for comparison.

@inproceedings{Barth2010Probabilistic,
  author = {Barth, Alexander and Siegemund, Jan and Mei{\ss}ner, Annemarie and Franke, Uwe and F\"orstner, Wolfgang},
  editor = {Goesele, M. and Roth, S. and Kuijper, A. and Schiele, B. and Schindler, K.},
  title = {Probabilistic Multi-Class Scene Flow Segmentation for Traffic Scenes},
  booktitle = {Pattern Recognition (Symposium of DAGM)},
  publisher = {Springer},
  year = {2010},
  pages = {503--512},
  note = {Darmstadt},
  doi = {10.1007/978-3-642-15986-2_51}
}

Alexander Barth and Uwe Franke, "Tracking Oncoming and Turning Vehicles at Intersections", In Intelligent Transportation Systems, IEEE Conference on. Madeira Island, Portugal, pp. 861-868. 2010.

• Abstract
• BibTeX
• DOI
• Download

This article addresses the reliable tracking of oncoming traffic at urban intersections from a moving platform with a stereo vision system. Both motion and depth information is combined to estimate the pose and motion parameters of an oncoming vehicle, including the yaw rate, by means of Kalman filtering. Vehicle tracking at intersections is particularly chal- lenging since vehicles can turn quickly. A single filter approach cannot cover the dynamic range of a vehicle sufficiently. We propose a real-time multi-filter approach for vehicle tracking at intersections. A gauge consistency criteria as well as a robust outlier detection method allow for dealing with sudden accelerations and self-occlusions during turn maneuvers. The system is evaluated both on synthetic and real-world data.

@inproceedings{Barth2010Tracking,
  author = {Barth, Alexander and Franke, Uwe},
  title = {Tracking Oncoming and Turning Vehicles at Intersections},
  booktitle = {Intelligent Transportation Systems, IEEE Conference on},
  year = {2010},
  pages = {861--868},
  doi = {10.1109/ITSC.2010.5624969}
}

Alexander Barth, "Vehicle Tracking and Motion Estimation Based on Stereo Vision Sequences". Thesis at: Institute of Photogrammetry, University of Bonn. 2010.

• Abstract
• BibTeX

Summary
In this dissertation, a novel approach for estimating trajectories of road vehicles such as cars, vans, or motorbikes, based on stereo image sequences is presented. Moving objects are detected and reliably tracked in real-time from within a moving car. The resulting information on the pose and motion state of other moving objects with respect to the own vehicle is an essential basis for future driver assistance and safety systems, e.g., for collision prediction The focus of this contribution is on oncoming traffic, while most existing work in the literature addresses tracking the lead vehicle. The overall approach is generic and scalable to a variety of traffic scenes including inner city, country road, and highway scenarios. A considerable part of this thesis addresses oncoming traffic at urban intersections. The parameters to be estimated include the 3D position and orientation of an object relative to the ego-vehicle, as well as the object's shape, dimension, velocity, acceleration and the rotational velocity (yaw rate). The key idea is to derive these parameters from a set of tracked 3D points on the object's surface, which are registered to a time-consistent object coordinate system, by means of an extended Kalman filter. Combining the rigid 3D point cloud model with the dynamic model of a vehicle is one main contribution of this thesis. Vehicle tracking at intersections requires covering a wide range of different object dynamics, since vehicles can turn quickly. Three different approaches for tracking objects during highly dynamic turn maneuvers up to extreme maneuvers such as skidding are presented and compared. These approaches allow for an online adaptation of the filter parameter values, overcoming manual parameter tuning depending on the dynamics of the tracked object in the scene. This is the second main contribution. Further issues include the introduction of two initialization methods, a robust outlier handling, a probabilistic approach for assigning new points to a tracked object, as well as mid-level fusion of the vision-based approach with a radar sensor. The overall system is systematically evaluated both on simulated and real-world data. The experimental results show the proposed system is able to accurately estimate the object pose and motion parameters in a variety of challenging situations, including night scenes, quick turn maneuvers, and partial occlusions. The limits of the system are also carefully investigated.
Zusammenfassung
In dieser Dissertation wird ein Ansatz zur Trajektorienschätzung von Straßenfahrzeugen (PKW, Lieferwagen, Motorräder, ... ) anhand von Stereo-Bildfolgen vorgestellt. Bewegte Objekte werden in Echtzeit aus einem fahrenden Auto heraus automatisch detektiert, vermessen und deren Bewegungszustand relativ zum eigenen Fahrzeug zuverlässig bestimmt. Die gewonnenen Informationen liefern einen entscheidenden Grundstein für zukünftige Fahrerassistenz- und Sicherheitssysteme im Automobilbereich, beispielsweise zur Kollisionsprädiktion. Während der Großteil der existierenden Literatur das Detektieren und Verfolgen vorausfahrender Fahrzeuge in Autobahnszenarien adressiert, setzt diese Arbeit einen Schwerpunkt auf den Gegenverkehr, speziell an städtischen Kreuzungen. Der Ansatz ist jedoch grundsätzlich generisch und skalierbar für eine Vielzahl an Verkehrssituationen (Innenstadt, Landstraße, Autobahn). Die zu schätzenden Parameter beinhalten die räumliche Lage des anderen Fahrzeugs relativ zum eigenen Fahrzeug, die Objekt-Geschwindigkeit und -Längsbeschleunigung, sowie die Rotationsgeschwindigkeit (Gierrate) des beobachteten Objektes. Zusätzlich werden die Objektabmaße sowie die Objektform rekonstruiert. Die Grundidee ist es, diese Parameter anhand der Transformation von beobachteten 3D Punkten, welche eine ortsfeste Position auf der Objektoberfläche besitzen, mittels eines rekursiven Schätzers (Kalman Filter) zu bestimmen. Ein wesentlicher Beitrag dieser Arbeit liegt in der Kombination des Starrkörpermodells der Punktewolke mit einem Fahrzeugbewegungsmodell. An Kreuzungen können sehr unterschiedliche Dynamiken auftreten, von einer Geradeausfahrt mit konstanter Geschwindigkeit bis hin zum raschen Abbiegen. Um eine manuelle Parameteradaption abhängig von der jeweiligen Szene zu vermeiden, werden drei verschiedene Ansätze zur automatisierten Anpassung der Filterparameter an die vorliegende Situation vorgestellt und verglichen. Dies stellt den zweiten Hauptbeitrag der Arbeit dar. Weitere wichtige Beiträge sind zwei alternative Initialisierungsmethoden, eine robuste Ausreißerbehandlung, ein probabilistischer Ansatz zur Zuordnung neuer Objektpunkte, sowie die Fusion des bildbasierten Verfahrens mit einem Radar-Sensor. Das Gesamtsystem wird im Rahmen dieser Arbeit systematisch anhand von simulierten und realen Straßenverkehrsszenen evaluiert. Die Ergebnisse zeigen, dass das vorgestellte Verfahren in der Lage ist, die unbekannten Objektparameter auch unter schwierigen Umgebungsbedingungen, beispielsweise bei Nacht, schnellen Abbiegemanövern oder unter Teilverdeckungen, sehr präzise zu schätzen. Die Grenzen des Systems werden ebenfalls sorgfältig untersucht.

@phdthesis{Barth2010Vehicle,
  author = {Barth, Alexander},
  title = {Vehicle Tracking and Motion Estimation Based on Stereo Vision Sequences},
  school = {Institute of Photogrammetry, University of Bonn},
  year = {2010}
}

Alexander Barth and Jan Siegemund and Uwe Franke and Wolfgang Förstner, "Simultaneous Estimation of Pose and Motion at Highly Dynamic Turn Maneuvers", In 31th Annual Symposium of the German Association for Pattern Recognition (DAGM). Denzler, J. and Notni, G. (Eds.) Jena, Germany, pp. 262-271. Springer. 2009.

• Abstract
• BibTeX
• DOI
• Download

Abstract. The (Extended) Kalman filter has been established as a stan- dard method for object tracking. While a constraining motion model stabilizes the tracking results given noisy measurements, it limits the ability to follow an object in non-modeled maneuvers. In the context of a stereo-vision based vehicle tracking approach, we propose and compare three different strategies to automatically adapt the dynamics of the fil- ter to the dynamics of the object. These strategies include an IMM-based multi-filter setup, an extension of the motion model considering higher order terms, as well as the adaptive parametrization of the filter vari- ances using an independent maximum likelihood estimator. For evalua- tion, various recorded real world trajectories and simulated maneuvers, including skidding, are used. The experimental results show significant improvements in the simultaneous estimation of pose and motion.

@inproceedings{Barth2009Simultaneous,
  author = {Barth, Alexander and Siegemund, Jan and Franke, Uwe and F\"orstner, Wolfgang},
  editor = {Denzler, J. and Notni, G.},
  title = {Simultaneous Estimation of Pose and Motion at Highly Dynamic Turn Maneuvers},
  booktitle = {31th Annual Symposium of the German Association for Pattern Recognition (DAGM)},
  publisher = {Springer},
  year = {2009},
  pages = {262--271},
  doi = {10.1007/978-3-642-03798-6_27}
}