Otto Loehlein

A cascade detector approach applied to vehicle occupant monitoring with an omni-directional camera

Recently Viola et al. have described a fast and robust face detection system using Haar-Wavelet-like features, AdaBoost and a classifier cascade. This paper deals with some handicaps of AdaBoost and proposes some modifications to Violas system. We then introduce a system for vehicle seat occupancy monitoring using an optical sensor.

Camera and imaging radar feature level sensorfusion for night vision pedestrian recognition

This contribution presents a robust pedestrian detection system at night that fuses a camera sensor and a scanning radar sensor on feature level. Each sensor defines an overdetermined set of features to be selected and parameterized using the supervised training algorithm AdaBoost. This technique assures an optimal selection and weighting of the features from both sensors depending on their discriminative power for the classification task. In the radar plane a new complex signal filter has been derived which describes a local similarity measure of velocity differences. In order to achieve realtime capability multiple classifiers are combined using a cascade.

Object management strategy for an unified high level automotive sensor fusion framework

An important requirement in autonomous driving for many complex scenarios is to correctly detect static and dynamic targets under various states of motion. The possibility of fulfilling this requirement depends upon the availability of different sensor data to the sensor fusion module. This paper uses data from sensors with built-in tracking modules and our objective is to make the resultant of two different sensor fusion modules that use the same sensor tracked data, to be statistically relevant based on the respective operational requirements despite this commmon prior set-up. In our case, we have two sensor fusion modules. One sensor fusion module deals with dynamic targets with well-defined object representation and other module deals only with static targets of undefined shapes. The authors have developed different concepts to manage the relevancy of the deliverables of the two modules. A novel approach based on multi-hypothesis tracking is presented. The results are evaluated using simulation and as well as with real world sensor data with reference ground truth target data.

Performance evaluation of FIR sensor systems applied to pedestrian detection

Besides resolution, an important performance parameter of a FIR camera is the sensitivity. It depends on the sensitivity of the detector array itself and the characteristics of the optic. The effects of the optic are considerably driven by the f-number, with high values resulting in decreased sensitivity, but providing the possibility for simple lens design and cheaper production costs. In this contribution 4 different sensor setups with different optics are evaluated for their impact on the performance of trained pedestrian classifiers. To overcome the expensive and time consuming process of ground truth generation for multiple sensors, an approach for reusing available high sensitivity reference data is presented. Classifiers are trained on specially transformed reference data with characteristics of sensors with degraded sensitivity. For the evaluation of the classifiers, data of real world road scenarios is collected simultaneously with the target sensors mounted in parallel in a test vehicle, following a detailed script for recording a pedestrian scene test catalogue. This allows for a direct analysis and comparison of the different sensors and their impact on the detection performance.

Sensor fusion for the detection of land mines

The detection of buried anti-personnel mines (APMs) is widely considered as a problem which may only be solved with a combination of two or more complementary sensors. We present processing and fusion results obtained from a multisensor data set, acquired with a pulse induction metal detector (MD), a pulsed ultra wide band ground penetrating radar (GPR) and a 3 - 5 (mu) thermal IR camera. Metal detector and GPR sensors were mounted on a rig for optimum control. Various types of soils, clutter objects and burial conditions were recorded. Anti personnel mines included minimum metal mines as well as mines with a significant metal content. We use a special projection to map a 3D GPR data cube, with time or depth as vertical coordinate, into a horizontal plane view 2D image. Object contours are then derived, based on an edge extraction method, followed by an automatic detection of circular shapes with a Hough-transform. In the association step, the stand-off IR image, the metal detector and GPR images and related detections are mapped onto a common cartesian grid on the ground surface. Detection results are fused on a decision level, using a Bayesian approach. Our results indicate that the GPR performance approximately matches that of the metal detector. With both sensors all metallic mines and around 60% of the minimum metal mines were detected. In the case of two false alarms per square meter combined detection probability clearly exceeds single sensor performance and reaches around 80%. Our fusion demonstrator, which incorporates all elements of the processing chain, has been implemented on the basis of MATLABTM.© (1999) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Detection of linear objects for synthetic vision applications

This paper describes an approach to find linear objects, e.g. powerlines and runways for synthetic vision applications with a multifunction 35 GHz radar. The approach is based on a combination of traditional radar signal processing like the CFAR-algorithm and image processing techniques like the Hough transform. It is assumed that the objects are visible as a sequence of single reflectors on a line. The proposed method ensures that the probability of detection or the false alarm rate of a linetype object is independent of the position. In the first step, a CFAR-algorithm detects the possible points along the line. All detected objects are describes by a list of attributes, from which some relevant ones can be chosen. Subsequently they are transformed to the Hough space, where lines are described by a slope and a distance parameter. A threshold is calculated which ensures a constant false alarm rate or a constant probability of detection. In the next step a cluster algorithm with a special distance measure is used to find all possible lines in the Hough-space. After transforming back to the original space, the plausibility is checked and a final selection is done. The performance of the approach is shown by applying the method described above to simulated and measured data. The paper describes the calculation of the false alarm rate, the probability of detection and the calculation of the threshold.