Efficient Methods for Video-Based Human Activity Analysis

Details

• Personal Author:
  Wang X
• Description:
  In this research project, we develop efficient Video Object Tracking (VOT) algorithms and video behavioral analysis algorithms for non-invasive observing and analyzing human activities. Two specific applications are investigated: monitoring driver distraction and ergonomics of factory workers performing heavy lifting jobs. We develop VOT algorithms to process a large scale SHRP-2 Naturalist Driving Study (NDS) dataset. The NDS dataset contains hundreds of thousands of hours of videos monitoring thousands of drivers driving on the road. The objective is to detect distractive driving events by processing billions of video frames. A key research effort is to develop a VOT algorithm that can significantly reduce the computing time without compromising the overall tracking performance. To this aim, we developed a novel temporal frame-subsampling (TFS) algorithm. Instead of processing every frame, the TFS algorithm performs video object tracking on a sub-sampled shorter video sequence obtained by sampling 1 frame out of N consecutive video frames from the original video sequence. The VOT tracking results (object's locations) of the N-1 frames between two consecutive sampled video frames will be obtained through interpolation of tracking results of the sampled frames. We show that for some NDS video sequences, N can be as large as 100 (100-fold speed-up) without sacrificing the VOT performance. The second application concerns body posture monitoring of industrial workers performing heavy lifting jobs. We developed a motion-segmentation based bounding box tracking algorithm to track the movement of a worker during the lifting operation. We show that the variations of the size of the bounding box can be used to infer the body posture (bending angle) of the worker during the lifting instant. Specific lifting and dropping events are detected by exploiting a ghosting phenomenon when the position of a lifted object changes from stationary to motion. This work validates the feasibility of using light-weight video analytics to estimate body posture without resorting to cumbersome and expensive 3D tracking devices. Along this direction, we also developed novel algorithms to estimate the body twisting (asymmetry) angle of the worker during a lifting event. The asymmetry angle is defined on the relative positions of the worker's wrists and ankles which are part of the skeletal joints of the worker's body. We incorporate state of art 2D and 3D skeletal joint estimation algorithm to aid the detection of 3D body joints locations. Then, we deduce the corresponding asymmetry angles. This algorithm is tested with a large dataset from a NIOSH laboratory. Very promising results have been observed. The works accomplished in this research provide strong evidence that a non-invasive video-based approach can provide accurate estimates of the states of human drivers during driving or workers during working. The algorithms developed in this work built on state-of-art computer vision algorithms. The main contributions of this work lie in the development of computation efficient formulation without losing performance. [Description provided by NIOSH]
• Subjects:
  Accidents, Occupational Automation Electrical Equipment And Supplies Mining Occupational Health Safety Safety Management Software
• Keywords:
  Analytical Instruments Computer Applications Drivers Ergonomics Exposure Assessment Lifting Monitoring Systems Posture Video
• ISBN:
  9781392687437
• Publisher:
  ProQuest LLC.
• Document Type:
  Text
• Funding:
  Grant
• Genre:
  Dissertation
• Place as Subject:
  OSHA Region 5 ; Wisconsin
• CIO:
  National Institute for Occupational Safety and Health (NIOSH)
• Topic:
  Workplace Safety & Health
• Location:
  North America
• NIOSHTIC Number:
  nn:20063351
• Citation:
  Ann Arbor, MI: ProQuest LLC., 2019 Aug; :27666535
• Federal Fiscal Year:
  2019
• Performing Organization:
  University of Wisconsin, Milwaukee
• Peer Reviewed:
  False
• Start Date:
  20160901
• Source Full Name:
  Efficient methods for video-based human activity analysis
• End Date:
  20190831
• Collection(s):
  National Institute for Occupational Safety and Health
• Main Document Checksum:
  urn:sha-512:d5c3ef01e244d487a150961e643b1e979d82701f570f5f5773ea5a8607b50d2f4e9962163893df40069124ab8d34bb531014d5b5aca83d79c1abd86332b58c3f
• Download URL:
  https://stacks.cdc.gov/view/cdc/226429/cdc_226429_DS1.pdf
• File Type:
  [PDF - 6.96 MB ]