 Telematics Applications for Transport Library
 Document item |
|
 Title : |
CROMATICA - TR 1016 - Developments of the new technologies |
| Author : |
S. Velastin; D. Aubert; J.L. Bruyelle; V. Caglioti; L. Khoudour; A. Langlais; M.A. Vicencio-Silva; M. Wherrett |
| Ref. No : |
TR1016 D9 |
| Date : |
20/08/97 |
| Size : |
96 pages |
Exec Summary : |
The addressed user needs have been grouped into Application Areas which group similar needs and/or required technology. Each Application Area has one or more demonstrators and a leading technical partner. The technologies necessary to address the user needs where identified and designed in the preceding stage (Deliverable D8).
The work in the phase reported here was concerned with the implementation and preliminary tests of modules to be integrated later in demonstrators. To ease development, each Application Area is relatively independent.
VP0: Video Library (KCL)
Most of the initial tests and a significant part of the demonstration phase, uses video recordings with examples of situations of interest to users. During this period, the number of video recordings has gradually increased. Legal problems in RATP sites (Paris) have been resolved and more extensive recordings are being made. Regular recordings in a large London Underground station will also commence shortly after appropriate interfaces have been made to the existing CCTV system.
VP1: Detection of potentially dangerous
situations in crowded conditions (KCL, UCL, PMI)
Three demonstrators have been defined:
a) Moderate-to-High crowding levels, good viewing perspective: Modules developed by KCL/UCL have been integrated into a complete working system ready for demonstration (for recordings or live video data). Preliminary tests have been carried out to evaluate consistency of results and detection performance. These have been successful but more recordings are required. A novel neural-based edge detector has been developed by PMI and is currently under validation. The main remaining work includes: ability to detect multiple situations simultaneously, possible integration of other partner's modules, migration to standard Windows environment.
b) Very high crowding levels, good viewing perspective: Novel algorithms based on local pixel intensity variability and neural-based classification have been developed and tested. Typical accuracy of discrimination of congestion levels (e.g. for six possible classes) is over 80% (greater than 90% for very crowded situations). The main remaining work includes improvement of classifier training stage, methods to deal with perspective distortion, optimisation of execution time. This method has potential to be used in other application areas.
c) Bad viewing perspective: No major work was planned for this stage.
VP2/VP3: Falls in the tracks, Intrusion Detection (INL, CAL)
Improvements to the STREAM image processor hardware are complete. Methods have been developed and evaluated to identify people and distinguish them from other smaller objects. The selection of a method suitable for live demonstration is under consideration. Preliminary tests have been conducted with real data obtained on-site. Techniques to distinguish between people and moving trains are being evaluated.
VP4: Abnormal motion and stationarity in corridors (IND)
Three demonstrators have been defined:
a) Abnormal stationarity: The most significant problem to deal with is that of occlusion (arising from perspective effects), specially for higher densities. The solution proposed consists in separating stationary and moving people with filtering to reduce noise and by delaying a decision when occlusion is detected. Preliminary tests have been successfully conducted, only limited by the amount of available video recordings. Subject to validation tests, this part is completed.
b) Fights: No significant work has been possible due to lack of data, but some possible approaches have been identified (e.g. measurement of patterns of motion).
c) Opposing flows: An optical flow technique has been selected to estimate directions of motion. Further processes have been added to improve the robustness of the measurements. These include spatial-temporal filtering, crowd trajectories tracking from image to image and flow segmentation into regions based in those trajectories. Preliminary tests have been conducted and some shortcomings identified (e.g. performance under low contrast or in the presence of regular patterns).
VP5: Level of occupancy (IND)
Two demonstrators have been defined:
a) Occupancy Level: From the previous stage the main problems identified were shadow removal and the need for a noise-filtering process. A morphological filtering process has been developed and tested. There has been less progress with shadow removals.
b) Queue measurements: Previous work in the project identified the main required technical developments: region extraction based on occupancy measurements, discrimation between queues and other stationary regions and shape analysis. The first two have been addressed, leaving shape analysis for the next stage. Preliminary tests have been conducted on the existing data and further work has been identified e.g. dealing with splitting/merging problems.
VP6: Routine data collection (MOL)
A scaleable architecture which is compatible and easily integrated with standard CCTV equipment has been developed and tested in situations with up to 1.1 million people per week being monitored by the system. Initial assessments of the system accuracy have been undertaken by comparison of the outputs of the system with the counts carried out by human observers over three hour periods. The variance of the routine data gathering system count compared to the human count over this period has typically been less than 5%.
The remainder of the development work during this work package will be focused on providing users with a versatile system in terms of the image algorithms that can be utilised and the format of the reports available to the operators.
This deliverable consists of 1 volume : cromatica d9.pdf |
|
