540 likes | 550 Views
Explore the use of machine vision to measure aerodynamic efficiency in heavy-haul railroads, evaluating gap lengths between loads for lower fuel costs. Discover challenges and proposed algorithms for improved efficiency.
E N D
Machine Vision Analysis of Intermodal Loading Efficiency on Heavy-Haul Railroads Tristan G. Rickett*, Avinash Kumar**, John M. Hart**, J. Riley Edwards*, Christopher P.L. Barkan* and Narendra Ahuja** *Rail Transportation & Engineering Center, University of Illinois at Urbana-Champaign, Urbana IL, USA; **Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, USA; IHHA 2011
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Goal : Measure Aerodynamic Efficiency • Lower fuel costs are incurred with aerodynamically efficient intermodal freight trains. • Aerodynamic efficiency can be estimated by computing all the gap lengths between successive loads of a freight train.[Lai, Barkan et.al, JRRT2006] • Machine vision provides us with efficient techniques to automatically detect gap lengths.[Kumar, Ahuja WACV06](This presentation)
Goal : Measure Aerodynamic Efficiency • Lower fuel costs are incurred with aerodynamically efficient intermodal freight trains. • Aerodynamic efficiency can be estimated by computing all the gap lengths between successive loads of a freight train.[Lai, Barkan et.al, JRRT2006] • Machine vision provides us with efficient techniques to automatically detect gap lengths.[Kumar, Ahuja WACV06](This presentation)
Goal : Measure Aerodynamic Efficiency • Lower fuel costs are incurred with aerodynamically efficient intermodal freight trains. • Aerodynamic efficiency can be estimated by computing all the gap lengths between successive loads of a freight train.[Lai, Barkan et.al, JRRT2006] • Machine vision provides us with efficient techniques to automatically detect gap lengths.[Kumar, Ahuja WACV06](This presentation)
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. • Cheap availability of high speed video cameras with good resolution (640x480). • Faster processing power of today’s computers. (Moore’s Law)
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. • Cheap availability of high speed video cameras with good resolution (640x480). • Faster processing power of today’s computers. (Moore’s Law)
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. • Cheap availability of high speed video cameras with good resolution (640x480). • Faster processing power of today’s computers. (Moore’s Law)
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. Accurate • Cheap availability of high speed video cameras with good resolution (640x480). • Faster processing power of today’s computers. (Moore’s Law).
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. Accurate • Cheap availability of high speed video cameras with good resolution (640x480). Cheap • Faster processing power of today’s computers. (Moore’s Law).
Machine Vision(MV) Based Approach • Advances in the field of computer vision and image processing techniques. Accurate • Cheap availability of high speed video cameras with good resolution (640x480). Cheap • Faster processing power of today’s computers. (Moore’s Law). Fast
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Challenges : Imaging Variations No Clouds Moving Clouds Varying Illumination Moving Trees
Challenges : Different Types of Loads • Empty Railcar • Double Stack with upper and lower stack of same length • & (d) Double Stack with upper and lower stack of different length • (e) Single Stack • (f) Trailer.
Challenges : Processing Speed • About 30 intermodal trains per day. • Each train has approx 5GB of information. • Processing huge data in limited time usually requires sacrificing some accuracy. • Find optimal balance between accuracy and time efficiency as both are important.
Challenges : Processing Speed • About 30 intermodal trains per day. • Each train has approx 5GB of information. • Processing huge data in limited time usually requires sacrificing some accuracy. • Find optimal balance between accuracy and time efficiency as both are important.
Challenges : Processing Speed • About 30 intermodal trains per day. • Each train has approx 5GB of information. • Processing huge data in limited time usually requires sacrificing some accuracy. • Find optimal balance between accuracy and time efficiency as both are important.
Challenges : Processing Speed • About 30 intermodal trains per day. • Each train has approx 5GB of information. • Processing huge data in limited time usually requires sacrificing some accuracy. • Find optimal balance between accuracy and time efficiency as both are important.
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
Example Video Acquisition@ Joliet, IL Viewing Volume Inter-modal Train Camera
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
TMS : Train Velocity Computation Successive Image Frames A B Find the location of image A in Image B
TMS : Train Velocity Computation Successive Image Frames Slide A over B by 0 pixel and compare 0 1
TMS : Train Velocity Computation Successive Image Frames Slide A over B by 1 pixel and compare 0 1
TMS : Train Velocity Computation Successive Image Frames Slide A over B by V pixel and compare V 0 1
TMS : Train Velocity Computation Successive Image Frames Best match obtained after shifting by V pixels V 0 1
TMS : Train Velocity Computation Successive Image Frames Train Velocity: V pixels/frame V 0 1
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
TMS : Background Subtraction • Background Subtraction is to separate the train from other objects • In train videos • locomotives/containers are the moving objects • sky/clouds/ground/trees/birds/… form background • Such natural (unrestricted) background is • diverse, and therefore • challenging to recognize and subtract
TMS : Background Subtraction • Background Subtraction is to separate the train from other objects • In train videos • locomotives/containers are the moving objects • sky/clouds/ground/trees/birds/… form background • Such natural (unrestricted) background is • diverse, and therefore • challenging to recognize and subtract
TMS : Background Subtraction • Background Subtraction is to separate the train from other objects • In train videos • locomotives/containers are the moving objects • sky/clouds/ground/trees/birds/… form background • Such natural (unrestricted) background is • diverse, and therefore • challenging to recognize and subtract
TMS : Background Subtraction x0 x0 x0+V Background Image Next Image Current Image A = Compare ~ 1 Proposed Foreground Measure : (A-B)/2 ~ 1 ~ -1 B = Compare -1 <= Compare( ) <= 1
TMS : Background Subtraction x1 x1 x1+V Background Image Next Image Current Image A = Compare ~ 1 Proposed Foreground Measure : (A-B)/2 ~ 0 ~ 1 B = Compare
TMS : Background Subtraction • Thus, foreground measure of 1 indicates train and 0 indicates background, • Applying this technique by selecting windows at the center of the image, background is removed.
TMS : Background Subtraction • Thus, foreground measure of 1 indicates train and 0 indicates background, • Applying this technique by selecting windows at the center of the image, background is removed. V pixels
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
TMS : Mosaic Generation • Repeating background removal for consecutive image frames leads to mosaic generation
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
TMS : Load Classification Double Stacks of two different kinds Trailer Single Stack
Flowchart of Machine Vision System Video Acquisition Train velocity computation Background removal Train Monitoring System (TMS) Mosaic Generation Load Classification Gap Length Estimation Train Scoring System (TSS) AEI Data Slot Efficiency & Aerodynamic Score
Outline • Goal • Machine Vision (MV) Based Approach • Challenges • Proposed MV Algorithms • Results Analysis • Conclusion & Future Work
Analysis of Gap Edge Accuracy I • 4 kinds of errors in gap estimation calculated • No Error: The boundary line detected by TMS correctly corresponds to the actual boundary of the load • Under Estimation: A small amount of background not removed near the load boundaries • Gap Not Detected: An extreme case of under estimation where the complete background between the loads is not removed • Over Estimation: A section of the load near the boundary is removed due to errors in background removal
Analysis of Gap Edge Accuracy II • Detailed analysis of ~ 1200 gaps