Karsten Kozempel

A comparison of methods for detecting atypical trajectories

The detection of atypical trajectories and events in road traffic is a challenging task for the implementation of an intelligent transportation system. It also provides information for optimizing the traffic flow and mitigating risks of accidents without the need to observe individual traffic participants. For detecting such events two methods representing the state of the art are compared: a map-based trajectory analyzer and a neural network, the Self Organizing Map—both applicable with unsupervised learning. The two compared algorithms detect atypical trajectories by modeling the probability function of trajectory features representing the object state at every trajectory point containing location, speed and acceleration values. The map-based approach was extended and improved by pre-clustering the trajectories with regard to their relation (e.g. vehicle turning left/going straight ahead). The Self Organizing Map algorithm uses vector quantization and prototyping of feature vectors and, thus, does not need any preliminary work. Both methods are evaluated by experiments using the same data which allows strengths and weaknesses to be revealed. The database for evaluation consists of trajectories from traffic surveillance cameras at an intersection and simulated trajectories.

Analysis of traffic safety for cyclists: the automatic detection of critical traffic situations for cyclists

Over the past years, the bicycle has gained importance as a means of transportation in big cities. The use and acceptance of a bicycle as being an evolving means of transportation is highly linked to its transportation safety. Still, the risk of accidents is a dominant barrier. Even though the Federal Ministry of Transport, Building and Urban Development established a National Cycling Plan to enhance cycling and improve safety aspects, serious accidents still occur. Even if the number of traffic accidents is declining in Berlin, the consequences of bicycle accidents with physical injury are characterised by increasing results. Thus, it is proved that more than half of the accidents that involve bicyclists are caused by the cyclist itself. To understand causes of accidents and to eventually arrange preventive measures and enhance cyclists’ safety, critical situations were detected. The application is based on cyclists’ trajectories generated from video sequences. As a result, atypical and dangerous traffic situations can be identified automatically whereas rule violations can be detected manually. First experiences at an intersection in Berlin show a general applicability of this approach, which has to be widely tested at other intersections.

Camera orientation based on matching road networks

The following paper shows an approach to measure orientation data of sensors during an airplane flight for traffic data acquisition. This orientation data is necessary for georegistration of any taken aerial images and extracted features. Since an inertial measurement unit is a very costly component, alternative methods for orientation determination should be considered. In this paper we present an approach achieving this only by using a GPS position and a database of the road network around. It is shown how the roads are extracted in the image considering their color hue and how they are matched to the given database segments. To adjust the rotation angles a simplex optimization algorithm is used. The approach yields convergence rates of more than 96 percent with a rotation error of around 1/4 degree for pitch and roll, which is sufficient for traffic surveillance purpose.

Road user tracker based on robust regression with GNC and preconditioning

In this paper an early vision tracking algorithm particularly adapted to the tracking of road users in video image sequences is presented. The algorithm is an enhanced version of the regression based motion estimator in Lucas-Kanade style. Robust regression algorithms work in the presence of outliers, while one distinct property of the proposed algorithm is that it can handle with datasets including 90% outliers. Robust regression involves finding the global minimum of a cost function, where the cost function measures if the motion model is conform with the measured data. The minimization task can be addressed with the graduated non convexity (GNC) heuristics. GNC is a scale space analysis of the cost function in parameter space. Although the approach is elegant and reasonable, several attempts to use GNC for solving robust regression tasks known from literature failed in the past. The main improvement of the proposed method compared with prior approaches is the use of a preconditioning technique to avoid GNC from getting stuck in a local minimum.

Wide-area based traffic situation detection at an ungated level crossing

The automated detection of atypical and critical traffic situations is essentially important to help to understand driver behaviour, to find functional correlations between traffic conflicts and real accidents, and eventually, to prevent, particularly severe accidents. In this paper a tool chain is introduced that enables a fully automated traffic situation detection in wide-area traffic on the basis of a single camera. The tool chain takes into account novel powerful methods for object detection, classification and tracking on the basis of robust regression with preconditioning on the one hand as well as traffic situation detection and classification on the basis of probabilistic approaches on the other hand and eventually, traffic event recording. The approach was tested at an ungated level crossing in the small town Bienrode, which is situated near Brunswick, Germany. It is shown that atypical situations, e.g. overtaking, braking, stopping, inadequate speeds and accelerations, as well as critical situations, e.g. tailgating, can be detected within a range of up to 120 m distance of the camera automatically. The approach enables new ways of analysing traffic areas with regard to traffic safety and performance. The results shown in this paper were obtained in the project OptiSiLK, whose abbreviation means “Optimisation of the safety and the performance at intersections of different traffic modes”. OptiSiLK was funded by the Ministry for Science and Culture of the State of Lower Saxony (MWK).

A Comparison of Trajectories and Vehicle Dynamics Acquired by High Precision GPS and Contemporary Methods of Digital Image Processing

Vehicle trajectories from recorded video sequences—acquired by several contemporary methods of digital image processing—are compared with high-precision GPS data serving as a reference. The raw data has been created by driving some scenarios with a car equipped with several sensors, i.e. Differential GPS (DGPS), acceleration sensor, etc. At the same time, the car was recorded by a video camera system in order to derive trajectory data by computer vision methods. Thus, the car is tracked by an Extended Kalman Filter (EKF) preceded by a background estimator. To improve the accuracy of the tracking data it is combined with a model-based approach for object detection. This approach fits a 3-dimensional wire frame model of the car into the image. The paper presents the driving scenarios of the car, the implemented image processing methods and a quantitative evaluation of the extracted trajectories obtained by two different image processing methods. Accuracy and precision of the methods are determined by comparing their results with the DGPS reference data of the car.

Airborne vehicle detection using SURF-descriptors and Support Vector Machines

Caused by the rising interest on traffic surveillance for simulations and decision management many publications focus on automatic vehicle detection systems. Vehicle counts and velocities of different car classes are the essential data basis for almost every traffic model. Especially during mass events or catastrophes conventional detection systems do not meet the demands. Thus a more flexible detector has to be used like an airborne camera system. In this paper a combination of a fast edge-based hypothesis generation and a more reliable hypothesis verification using a Support Vector Machine is presented. Due to image sizes of more than 20 megapixels at first the region of interest has to be preselected using a street database. Afterwards the first detection stage of the algorithm generates object hypotheses using especially shaped edge filters. The second detection stage verifies them by extracting the SURF-descriptor of each hypothesis. A Support Vector Machine is used to decide whether the objects descriptor represents a vehicle. It will be shown how the verification stage improves the detection reliability by discarding false positives while preselection and hypothesis generation provides less computation time.