Joint Human Detection from On-Board
        and Off-Board Cameras
       Justinas Mišeikis and Paulo V. K. Borges
Problem
Various vehicles and people share the same environment.




   Unfortunately, in these conditions accidents occur.
Problem


Can we use technology to
     prevent them?
Related Work
• Pedestrian detection using vehicle mounted
  stereo cameras
• Crowd tracking in enclosed environments
• People tracking using fixed cameras and
  driver warning if a person is on the path
• Same object identification using off-board
  cameras and not localized mobile camera
Our Method Overview
   Off-Board camera   In world coordinates




    Pos relative to   On-Board camera
     the vehicle         detection
Our Method Overview
Off-Board Cameras
• Fixed cameras
• MOG2 background segmentation
• Noise filtering
• Blob detection
• Size and dimensions filtering
• Feet position - bottom centre
  point of the blob
People Detector - HOG
• Popular method for people detection
• Works well in cluttered environments
• Based on distribution of intensity gradients
  or edge directions.
• Descriptor created from many samples
On-Board Cameras
           Full Image Analysis




The camera view is split according    A: Whole Image analysis - person not
 to the distance from the camera:                    detected
  blue (3-7 meters), green (7-12     B: Area split method - person detected
  meters), yellow (12-20 meters).
Targeted Analysis

Expected feet                                Detected
position from                                 person
ground plane
 homography



                  HOG search area
  Size calculated using the expected person height
Data Fusion
    Off-Board Cam A              Off-Board Cam B


Detected      On-Board Cam - HOG
 people
positions
                  Position Fusion


              Final Pedestrian Position
Data Fusion
    Position sensor fusion

                                                                   6



    HOG detector variance



                               Estimated distance error (meters)
                                                                            Measurement Error
                                                                   5
                                                                            Polynomial Error Estimation



    estimation depending                                           4




    on the object’s distance
                                                                   3



                                                                   2



    from the camera                                                1



                                                                   0
                                                                    2   3       4     5     6     7       8   9   10   11   12
                                                                             Distance from the camera (meters)




Extended Kalman Filter for sensor fusion and tracking
Implementation
• C++
• Standard Intel laptop + desktop running in
  parallel
• cvBlob library
Experiments
    30m by 30m industrial site
    Test vehicle - HMC
    2 Off-Board, 1 On-Board cam
    Three 4-6 minute runs
    1-2 People walking around
    Vehicle static and moving
Results
                        FP %     FN %
  Localized Analysis     4.41     3.66
  Full Image Analysis   95.53     2.40


• Works real-time
• Background Segmentation - 10-15 FPS
• HOG detectors - 5 FPS
Progress
• Additional front facing cameras, 4 in total
 - Two side cameras
 - Wide angle camera for close proximity
 - Far proximity front camera
• People tracking
• HOG performed on GPU GeForce GT640
Progress
                                  Close Prox Cam




  Danger Zone -
reduced to around
  20 cm from the
front of the vehicle
                                   Far Prox Cam
Thank You!

Any Questions?

Joint Human Detection from On-Board and Off-Board Cameras

  • 1.
    Joint Human Detectionfrom On-Board and Off-Board Cameras Justinas Mišeikis and Paulo V. K. Borges
  • 2.
    Problem Various vehicles andpeople share the same environment. Unfortunately, in these conditions accidents occur.
  • 3.
    Problem Can we usetechnology to prevent them?
  • 4.
    Related Work • Pedestriandetection using vehicle mounted stereo cameras • Crowd tracking in enclosed environments • People tracking using fixed cameras and driver warning if a person is on the path • Same object identification using off-board cameras and not localized mobile camera
  • 5.
    Our Method Overview Off-Board camera In world coordinates Pos relative to On-Board camera the vehicle detection
  • 6.
  • 7.
    Off-Board Cameras • Fixedcameras • MOG2 background segmentation • Noise filtering • Blob detection • Size and dimensions filtering • Feet position - bottom centre point of the blob
  • 8.
    People Detector -HOG • Popular method for people detection • Works well in cluttered environments • Based on distribution of intensity gradients or edge directions. • Descriptor created from many samples
  • 9.
    On-Board Cameras Full Image Analysis The camera view is split according A: Whole Image analysis - person not to the distance from the camera: detected blue (3-7 meters), green (7-12 B: Area split method - person detected meters), yellow (12-20 meters).
  • 10.
    Targeted Analysis Expected feet Detected position from person ground plane homography HOG search area Size calculated using the expected person height
  • 11.
    Data Fusion Off-Board Cam A Off-Board Cam B Detected On-Board Cam - HOG people positions Position Fusion Final Pedestrian Position
  • 12.
    Data Fusion Position sensor fusion 6 HOG detector variance Estimated distance error (meters) Measurement Error 5 Polynomial Error Estimation estimation depending 4 on the object’s distance 3 2 from the camera 1 0 2 3 4 5 6 7 8 9 10 11 12 Distance from the camera (meters) Extended Kalman Filter for sensor fusion and tracking
  • 13.
    Implementation • C++ • StandardIntel laptop + desktop running in parallel • cvBlob library
  • 14.
    Experiments 30m by 30m industrial site Test vehicle - HMC 2 Off-Board, 1 On-Board cam Three 4-6 minute runs 1-2 People walking around Vehicle static and moving
  • 15.
    Results FP % FN % Localized Analysis 4.41 3.66 Full Image Analysis 95.53 2.40 • Works real-time • Background Segmentation - 10-15 FPS • HOG detectors - 5 FPS
  • 16.
    Progress • Additional frontfacing cameras, 4 in total - Two side cameras - Wide angle camera for close proximity - Far proximity front camera • People tracking • HOG performed on GPU GeForce GT640
  • 17.
    Progress Close Prox Cam Danger Zone - reduced to around 20 cm from the front of the vehicle Far Prox Cam
  • 18.