Real-time quality assessment of videos from body-worn cameras

Abstract

 

Videos captured with body-worn cameras may be affected by distortions such as motion blur, overexposure and reduced contrast. Automated video quality assessment is therefore important prior to auto-tagging, event or object recognition, or automated editing. In this paper, we present M-BRISQUE, a spatial quality evaluator that combines, in realtime, the Michelson contrast with features from the Blind/Referenceless Image Spatial QUality Evaluator. To link the resulting quality score to human judgement, we train a Support Vector Regressor with Radial Basis Function kernel on the Computational and Subjective Image Quality database. We show an example of application of M-BRISQUE in automatic editing of multicamera content using relative view quality, and validate its predictive performance with a subjective evaluation and two public datasets.

 

 

 

Application to multi-view camera selection

 

 

 

 

Real-time demo

 

 

 

 

Reference

 

Real-time quality assessment of videos from body-worn cameras
     Yuan-Yi Chang, Riccardo Mazzon, Andrea Cavallaro
     EUSIPCO 2018 (to appear)