Pedestrian Re-identification Based On Feature Fusion And Metric Learning

Pedestrian re-identification is the problem of identifying the same person appeared in different camera views.It is an important problem that has gained more and more interests in recent years,which has high potential applications in areas like visual surveillance,person tracking and image retrieval.Pedestrianre-identification still remains a challenging task due to difficulties like influence of occlusions,pose changes,different lighting conditions and low resolutions.The main contents of our research work include:(1)We designed a pedestrian re-identification flowchartincluding three main steps: pre-processing which includes pedestrian segmentation and body part subdivision,feature extraction and pedestrian matching.(2)We proposed a novel fusion based feature representationthat combines high-dimension low-level features(color histograms and HOG features)and low-dimension mid-level descriptors(color name descriptor);(3)We built a new pedestrian re-identification dataset named as CampusPed which was more close to real environment and more complicated with increasingnumber of pedestrians;(4)Based on feature fusion strategy we adopted a metric learning algorithm KISSME to do person matching.Experiments on several public pedestrian re-identification datasets(VIPeR,i-LIDS and CAVIAR4REID)demonstrate that our approach achieves state-of-the-art results compared to other known published approaches.The top-1 results are 95.25%,50.85%,69.44% respectively,which are significantly better than other known methods.