pattern recognition presentation
TRANSCRIPT
Illumination Assessment for Vision-Based Traffic Monitoring
BySHRUTHI KOMAL GUDIPATI
Outline
Introduction PVS system design & concepts Assessing lighting Assessing contrast Assessing shadow presence Conclusion
Introduction Vision systems in traffic domain operates
autonomously over varying environmental conditions
Uses different parameter values or algorithms depending on these conditions
Parameters depends on ambient conditions on camera images
PVS system Commercial real-time vision system for traffic
monitoring that detects, tracks, and counts vehicles
Uses large volume of video data obtained from 25 different scenes
Switches between different parameter values and algorithms depending on scene illumination aspects
Aspects of Scene illumination
Is the scene well-lit? Is vehicle bodies visible? In poorly lit scenes, Are only vehicle
lights visible ?
Aspects of Scene illumination
Aspects of Scene illumination
Aspects of Scene Iillumination
Is the contrast sharp enough? Ex:
Is visibility sufficient for reliable detection ?
Is visibility sufficient or too diminished ?
Fog, Dust or Snow
Aspects of Scene illumination
Are vehicles in the scene casting shadows?
PVS System Design
Processes frames at 30 Hz
Process images simultaneously up to 4 cameras
Compact and fits in 3U VME board
3U VME Board
PVS system hardware
Two Texas Instruments TMS320C31 DSP chips
A Sensar pyramid chip Custom ALU implemented using a
Xilinx chip
Operation principle Maintains a reference image that contains
the scene as it would appear if no vehicles were present
Each incoming frame is compared to the reference
Pixels where there are significant differences are grouped together into "fragments" by the detection algorithm
These fragments are grouped and tracked from frame to frame using a predictive filter
One dimensional strip representation
Reduces the 2D image of each lane to a 1D "strip“
Integration operation that sums two pixel-wise measures across the portion of each image row that is spanned by the lane, resulting in a brightness and energy measurement for each row
Integration operation is performed by the ALU, which takes as input a bit-mask identifying each lane
2D -> 1D transformation
Strip measurements Two measurements, brightness and energy,
are computed for each strip element y of each strip s
Brightness B(s,y) = Σ (pixels inWy)
Energy E(s,y) = [Σ (absolute difference between every two
adjacent pixels in W) ] / ااWyاا
Reference strips
Brightness and Energy measurements gathered from a strip over time are used to construct a reference strip
For scenes in which traffic is flowing freely, reference strip can be constructed by IIR filtering
IIR filtering doesn’t work in stop-and-go or very crowded areas
Strip element classification Classify each strip element on the
current strip as background or non-background
Done by computing the brightness and energy difference measures ΔB(y) = B(I,y) - B (R,y) - (o اا W(y) اا ) ΔE(y) = اE(I,y) - E(R,y) ا
Classification as Background or non-Background
Strip element classification
Each strip element that is classified as non-background is further classified as "bright" or "dark“
Depends on whether its brightness is greater or less than the brightness of the corresponding reference strip element
Illumination Assessment
All frames grabbed in a two-minute interval, all strip elements that both have been identified as non- background and have significant dt are used to update various statistical measures
Values of these measures are used to assess the lighting, contrast, and shadows
Fragment Detection
Groups non-background strip elements into symbolic "vehicle fragments“
To prevent false positive vehicle detections, the system avoids detecting the illumination artifacts as vehicle fragments
Fragment Detection
Uses three different detection techniques, depending on the nature of the scene illumination Detection in well-lit scenes without vehicle
shadows Detection in well-lit scenes with vehicle shadows Detection in poorly-lit scenes
Fragment Detection
Detection in well-lit scenes without vehicle shadows
Scene as well-lit if the entire vehicle body is Visible
Scenes are termed poorly-lit if the only clearly-visible vehicle components are the headlights or taillights
Fragment Detection
Detection in well-lit scenes with vehicle shadows
Well-lit scenes where vehicles are casting shadows, the detection process must be modified so that non-background strip elements due to shadows are not grouped into vehicle fragments
Uses stereo or motion cues to infer height
Fragment Detection
Detection in poorly-lit scenes
Where only vehicle lights are visible, fragment extraction via connected components is prone to false positives due to headlight reflections
Fragments are extracted by identifying compact bright regions of non-background strip elements around local brightness maxima
Fragment tracking & grouping
After the vehicle fragments have been extracted, they are passed to the Tracker module which tracks over time and groups them into objects
Assessing lighting Measures used for assessing whether the scene is
well-lit, i.e. whether the entire body of most vehicles will be visible Ndark + Nbright = total number of non-
background pixels that were detected Pdark = Ndark/(Ndark+Nbright)
If the scene is poorly-lit, the background image will be quite dark, and it will be difficult to detect any pixel with a dark surface color. Under this condition ndark will be small, and hence Pdark will be small
Assessing contrast
Two typical causes of insufficient contrast -- fog or raindrops
Contrast can be measured using the energy difference measure ΔE(y)
In low-contrast scenes that occur during the day, vehicles will usually appear as objects darker than the haze, which often appears rather bright
In low-contrast scenes occurring at night, no dark regions will be detectable
Measure ΔEbright and ΔEdark
Assessing shadow presence
Scenes that are well-lit can be decomposed into two sub-classes Shadows Non - Shadows
Contrast of a "bright" portion of a vehicle against the road surface would be less than that of a "dark" portion
Assessing shadow presence
Using k4 = 1.2, this method has been found to work well
Sometimes, when there are very faint shadows, it does classify the scene as having no shadows
Fails when the background is not a road For example, in some scenes a camera is
looking at the road primarily from the side, and the vehicles occlude either objects (e.g. trees) or the sky as they move across the scene
Illumination Assessment module
Three methods for assessing lighting, contrast, and shadows are applied sequentially
Illumination Assessment module
Conclusions
During Strip representation, transformation from 2D -> 1D is not clearly explained
In strip classification, the global offset “o” is mentioned to have been measured by a different process. The paper doesn’t mention/explain anything about the process
The paper mentions that the deployment results were satisfactory but it doesn’t provide any statistical data to support the claim
References
Wixson, L.B., Hanna, K., Mishra, D., Improved Illumination Assessment for Vision-Based Traffic Monitoring, VS98(Image Processing for Visual Surveillance)
Hanna, K. L. Wixso and D. Mishra , Illumination Assessment for Vision-Based Traffic Monitoring, ICPR '96: Proceedings of the International Conference on Pattern Recognition
Femer et al. 941 N.J. Ferrier, S.M. Rowe, A. Blake, "Real-Time Traffic Monitoring," in Proceedings of the IEEE Workshop on Applications of Computer Vision, pages 81-88, 1994
Kilger 911 M. Kilger, "A Shadow Handler in a Video-based Real-time Traffic Monitoring System", in Proceedings of the IEEE Workshop on Applications of Computer Vision, pages 11-18, 1992
Questions ?