Sparse Appearance Model In Infrared Pedestrian Tracking

At present,video surveillance,human-computer interaction,assisted driving,sports broadcast and so on,all need pedestrian tracking.It is one of the most important field in computer vision.Thermal infrared imaging does not rely on external illumination,provides the possibility for day and night differential applications.Pedestrian tracking on thermal infrared imaging has high research significance and wide application value,but the low contrast,excessive noise,less target information,pedestrian posture diversity and occlusion make it difficult to work.At present,the representation of over-complete base in sparse coding makes the tracking in thermal infrared image very effective,but most of the algorithms are for general objects,and ignore the impact of pedestrian posture changes,so it lacks of adaptability in the real scene,and cannot achieve accurate tracking;on the other hand,the existing tracking algorithm is generally focused on how to deal with the target is blocked or other changes in the appearance of the situation,once the algorithm lost its effectiveness,it will be irreducible to produce serious drift problems.In the realistic scenes,occlusion and appearance changes inevitable,and the robustness of tracking must also decline,but the more important thing is how can we recover and keep tracking accurately after that.In respect of the issues above,this paper deeply studies sparse coding and the pedestrian structural information in infrared light,proposes the local sparse appearance model with structural information and combine it with the confidence level evaluation to track.Firstly,the candidates are divided into three parts according to the pedestrian’s variation degree commonly in infrared image sequence,then a set of local image patches are extracted from each part in templates as dictionary.Secondly,we take two kinds of sparse coding to represent the target vector.One is local sparse appearance model,which uses two-step pooling method to adapt the moving characteristics of pedestrian.The other is locality-constrained sparse appearance model,which only uses one-step pooling to meet the pedestrians’ nonrigid requirements and accelerate the calculation.Thirdly,to handle the scene of appearance changes drastically,the confidence level evaluation is used and a new template update selecting strategy is exploit on it.In the experiment,we evaluate the proposed algorithm on several representative infrared image sequences and OSU dataset,compare it with other six state-of-the-art algorithm and demonstrate that the proposed algorithm can outperform the other start-of-the-art algorithms when occluding,scale changing,cluster background occurs and high speed.