Extended Observation Particle Filter with SVD Template Generation Implemented for GPU

Abstract

This work presents a novel template updating strategy based on singular value decomposition (SVD) together with an expansion and extension of previous work combining observations across temporally adjacent frames to implement a likelihood function that provides improvement to velocity refinement in a particle filter tracker. SVD as a novel approach to template generation is used to take advantage of the intuitive notion that the largest singular value corresponds to the highest correlate across template candidates which should more adequately represent the target appearance while rejecting noise and other distractions for use in a correlation based scoring system such as the proposed likelihood function extended across temporally adjacent frames. The tracker is implemented to accelerate computationally expensive operations by moving them to the GPU for processing. This proposed expanded likelihood function provides an improvement of 11.2%-12.5% to particle degeneracy as compared to the previous method in the "Augmented State Vector" approach across the composite of videos. This improvement to particle degeneracy provides for a lower requirement in the number of actual particles necessary for implementation of a particle filter tracker and thus a lower computational requirement while simultaneously providing similar performance. The proposed SVD template generation provides 23.8% increase in time before track loss when comparing the best case in each category of update-by-score and update-by-SVD across the composite of videos. While not bench-marked in this work for quantitative comparison, the use of the GPU with Tensorflow-GPU and Python has allowed the large data set needed for analysis to be obtained in days instead of the months that would have been required in my original proof-of-concept work that was targeted solely for CPU in Matlab.