Traffic Information Extraction from Visual Sensors
Many valuable parameters describing the behaviour of multiple objects in a complex scene can be obtained with single-sensor systems only [1,2]. Using a traffic surveillance camera, background estimation, feature extraction and object extraction can be performed to derive object state vectors describing the positions and behaviour of detected targets.Still these systems are susceptible to changes of lighting, weather, occlusions and in the majority of cases do not perform well in congestion situations.
Multi-Sensor Settings
The use of multi-sensor networks can allow for better reliability and accuracy of measurements. Depending on the number of overlapping sensors, faulty data can be identified and ignored. In this work, a homogenous, cooperative and overlapping network configuration is used.
Multi-Sensor Data Fusion
Because of overlapping sensor ranges, the same parameters are measured multiple times. The deviating and occasionally conflicting information has to be processed with an appropriate data fusion approach, e.g. statistic fusion, voting fusion, bayesian decision fusion or dempster-shafer approaches.
The graph shows hypothetical class-conditional probability densities, the deviating probabilities to measure particular values if it is already known that the target to measure falls into one of two categories. Bayes rule applies.
With a dempster-shafer approach, counter-proof can be used to change the plausibility of uncertainty intervals for a particular measurement.
Recent advances in vision-based tracking systems lead to growing possibilities for the derivation of abstract scene descriptions and automatically obtained situation awareness. The thesis and related work is concerned with the analysis and interpretation of data supplied by multiple sensors observing the same scene or multiple scenes that are spatially overlapping or closely related.
Processing of data from sensor level to object information level
The figure above depicts a typical processing pipeline which does a signal adjustment, background estimation, feature extraction (e.g. corners, lines, vehicle contours) and the segmentation of the image which eventually results in object data in image coordinates.
Image coordinates are transformed into world space by application of an inverse projection matrix accounting for the intrinsic and extrinsic calibration matrix of the camera.
Multi-Sensor setup with decentralized feature extraction
The processing pipeline can be realized decentralized with modern surveillance sensors. Each sensor does the processing on its own, generating object data from sensor level data. Fusion of the resulting object state vectors can therefor be approached at object data level or even at a higher information or decision level.
Still this information shows high dimensionality and has to be reduced to significant dimensions to come to an understanding of a complex traffic situation (see 'Derivation of abstract traffic parameters')
Detection of atypical Events
With higher order traffic parameters generated by multi-sensor networks, detection of atypical or dangerous data constellations in complex situations can be realized. Atypical levels of load for particular lanes at rush hour or rogue driving detection can be examples for next generation traffic parameters.
Modeling of Object Data
By projecting image coordinates in the ground plane of the traffic situation, object coordinates of vehicles and other targets are available in world coordinates.
The application of a tracking algorithm (e.g. Kalman filtering, Multi-Hypothesis Tracking, Particle Filtering) yields a data array containing object positions attributed to targets moving through the scene.
Derivation of abstract traffic parameters
Object state vectors given, mean position (active areas), velocities and accelerations can be computed. Higher order data results from e.g. evaluating acceleration at a particular velocity or velocities together with the position of an object (e.g. atypically moving fast in a curve or very slowly along a straight road).
Active area and hough parameter space for lane extraction
Next order parameters include amongst others lane extraction, the clustering of entry- and exit-locations in the scene, average velocities for traffic lanes, mean vehicle separation and number of lane shifts.
Visualization for Situation Awareness
To contribute to the traffic safety and support incident management procedures results of traffic analysis can be visualized.
Human situation awareness may be increased by showing data of reduced complexity and emphasizing views of high significance.
References
[1] D. Beymer et al., A real-time computer vision system for measuring traffic parameters. 1997.[2] Koller et al. Robust multiple car tracking with occlusion reasoning. In ECCV (1), 189-196, 1994.