Robust infrared vehicle tracking across target pose change using L1 regularization

In this paper, we propose a robust vehicle tracker for Infrared (IR) videos motivated by the recent advance in compressive sensing (CS). The new eL₁-PF tracker solves a sparse model representation of moving targets via L₁ regularized least squares. The sparse-model solution addresses real-world environmental challenges such as image noises and partial occlusions. To further improve tracking performance for frame-to-frame sequences involving large target pose changes, two extensions to the original L₁ tracker are introduced (eL₁). First, in the particle filter (PF) framework, pose information is explicitly modelled into the state space which significantly improves the effectiveness of particle sampling and propagation. Second, a probabilistic template update scheme is designed, which helps alleviating drift caused by a target pose change. The proposed tracker, named eL₁-PF tracker, is tested on IR sequences from the DARPA Video Verification of Identity (VIVID) dataset. Promising results from the eL₁-PF tracker are observed in these experiments in comparison with previous mean-shift and original L₁-regularization trackers.