CONCLUSION & FUTURE WORK VEHICLE DETECTION IMAGE PROCESSING VISTA – COMPUTER VISION INNOVATIONS FOR SAFE TRAFFIC VEHICLE ORIGIN DETECTION USING LICENSE PLATE RECOGNITION OPTIMIZATION AND SPEED UP Overall objectives: •Strengthening of technology transfer and commercialization capacities of partner HEIs •Transfer of existing computer vision applications from HEIs to SMEs •Developing new traffic- and transportation-related computer vision applications with commercial potential in collaboration with SMEs in the automotive industry sector Measurement of road traffic parameters based on computer vision ACKNOWLEDGEMENTS Basic work flow of vehicle detection consists of: •Original image (a) down sampled and blurred (b) •Creation of background image model and comparison of background image model and current image (c) •Pixel clusterization and grouping (d) VEHICLE TRACKING AND COUNTING •In the proposed system, spatio-temporal tracking of objects in a scene is used for filtering •Every currently tracked object in the scene is compared with each cluster detected in the current image •When a vehicle passes through marker and a hit is detected, counter for that marker is incremented University of Zagreb, Faculty of Transport and Traffic Sciences www.unizg.hr, www.fpz.unizg.hr, its.fpz.hr •Faculty of Electrical Engineering and Computing (UNIZG-FER) leading institution •Project leader: prof. dr. sc. Sven Lončarić •Faculty of Transport and Traffic Sciences (UNIZG-FTTS) partner institution •Local project leader: prof. dr. sc. Hrvoje Gold •Project start date / duration: 30.03.2013. / 24 months •Project budget: 685,265.45 € This work has been supported by the IPA2007/HR/16IPO/001-040514 project “VISTA - Computer Vision Innovations for Safe Traffic” which is co- financed by the European Union from the European Regional and Development Fund. a) b) c) d) Kristian Kovačić Edouard Ivanjko Niko Jelušić Hrvoje Gold kkovacic@fp z.hr eivanjko@fpz .hr njelusic@f pz.hr hgold@fpz .hr Development of: •Detection of road-side vegetation for traffic infrastructure maintenance •Surround-view parking visualization •Traffic sign detection and recognition •Lane detection and recognition •Lane departure and collision warning •Automatic headlight detection •Driver mental state recognition Approach Vehicle count Tot al Left Righ t Overlap check Hits 126 65 61 FP / FN 0 / 6 0 / 5 0 / 1 Accuracy [%] 95, 6 92,9 98,4 Trajecto ry check Hits 129 68 61 FP / FN 1 / 4 0 / 3 1 / 1 Accuracy [%] 96, 2 95,8 96,8 Groundtruth 132 70 62 1. Video input 2. Using contours for vehicle detection 3. Variable image size for License Plate Recognition Country Number of vehicle s Percent age [%] Germany 166 31.2 Poland 88 16.5 Austria 83 15.6 Czech Republic 72 13.5 Croatia 47 8.8 Others 76 14.4 4. Vehicle country of origin distribution table •Developed system uses only one camera to detect and track vehicles in real time on a road with multiple lanes •First testing results are promising with vehicle detection accuracy of over 95% •Future work consists of developing a multiple object tracking system which would perform vehicle classification and therefore separate vehicles by their type, compute origin- destination matrices by identifying each individual vehicle and tracking it through monitored road traffic network Computed error between estimated (EKF) and real trajectory Counted vehicles in a video footage from Croatian highway near Zagreb Execution time needed for processing of a single image Architecture of optimized application 22 ND SUMMER SCHOOL ON IMAGE PROCESSING – SSIP 2014, ZAGREB, CROATIA

CONCLUSION & FUTURE WORK

Download PPTX Report

Upload
lancelot-marin
View
23
Download
0

Tags:

Embed Size (px)

DESCRIPTION

Measurement of road traffic parameters based on computer vision. University of Zagreb, Faculty of Transport and Traffic Sciences www.unizg.hr, www.fpz.unizg.hr, its.fpz.hr. VISTA – COMPUTER VISION INNOVATIONS FOR SAFE TRAFFIC. Development of: - PowerPoint PPT Presentation

Citation preview

CONCLUSION & FUTURE WORK

VEHICLE DETECTIONIMAGE PROCESSING

VISTA – COMPUTER VISION INNOVATIONS FOR SAFE TRAFFIC

VEHICLE ORIGIN DETECTION USINGLICENSE PLATE RECOGNITION

OPTIMIZATION AND SPEED UP

Overall objectives:• Strengthening of technology transfer and commercialization capacities of partner HEIs

• Transfer of existing computer vision applications from HEIs to SMEs

• Developing new traffic- and transportation-related computer vision applications with commercial potential in collaboration with SMEs in the automotive industry sector

Measurement of road traffic parametersbased on computer vision

ACKNOWLEDGEMENTS

Basic work flow of vehicle detection consists of:

• Original image (a)down sampled and blurred (b)

• Creation of background imagemodel and comparison ofbackground image model andcurrent image (c)

• Pixel clusterization andgrouping (d)

VEHICLE TRACKING AND COUNTING• In the proposed system, spatio-temporal tracking of objects in a sceneis used for filtering

• Every currently trackedobject in the scene iscompared with eachcluster detected in thecurrent image

• When a vehicle passesthrough marker and a hitis detected, counter forthat marker is incremented

University of Zagreb, Faculty of Transport and Traffic Scienceswww.unizg.hr, www.fpz.unizg.hr, its.fpz.hr

• Faculty of Electrical Engineering andComputing (UNIZG-FER) leading institution

• Project leader: prof. dr. sc. Sven Lončarić• Faculty of Transport and Traffic Sciences (UNIZG-FTTS) partner institution

• Local project leader: prof. dr. sc. Hrvoje Gold• Project start date / duration: 30.03.2013. / 24 months• Project budget: 685,265.45 €

This work has been supported by the IPA2007/HR/16IPO/001-040514 project “VISTA - Computer Vision Innovations for Safe Traffic” which is co-financed by the European Union from the European Regional and Development Fund.

a) b)

c)d)

Kristian Kovačić Edouard Ivanjko Niko Jelušić Hrvoje [email protected] [email protected] [email protected] [email protected]

Development of:• Detection of road-side vegetation for traffic infrastructure maintenance

• Surround-view parking visualization

• Traffic sign detection and recognition

• Lane detection and recognition

• Lane departure and collision warning

• Automatic headlight detection

• Driver mental state recognition

ApproachVehicle count

Total Left Right

Overlap check

Hits 126 65 61FP / FN 0 / 6 0 / 5 0 / 1

Accuracy [%] 95,6 92,9 98,4

Trajectory check

Hits 129 68 61FP / FN 1 / 4 0 / 3 1 / 1

Accuracy [%] 96,2 95,8 96,8Groundtruth 132 70 62

1. Video input 2. Using contours for vehicle detection 3. Variable image size for

License Plate Recognition

Country Number of vehicles

Percentage [%]

Germany 166 31.2Poland 88 16.5Austria 83 15.6Czech Republic 72 13.5Croatia 47 8.8Others 76 14.4

4. Vehicle countryof origin

distribution table

• Developed system uses only one camera to detect and track vehicles in real time on a road with multiple lanes

• First testing results are promising with vehicle detection accuracy of over 95%

• Future work consists of developing a multiple object tracking system which would perform vehicle classification and therefore separate vehicles by their type, compute origin-destination matrices by identifying each individual vehicle and tracking it through monitored road traffic network

Computed error between estimated (EKF) and real trajectory

Counted vehicles in a video footagefrom Croatian highway near Zagreb

Execution time needed for processing of a single image

Architecture of optimized application

22ND SUMMER SCHOOL ON IMAGE PROCESSING – SSIP 2014, ZAGREB, CROATIA