TOP  >  Thesis/Dissertation  >  {sotsurontitle}

Inter-Camera Person Re-Identification Using Weighted Features for Each Pair of Cameras

In this thesis, we present an appearance based method for inter-camera person re-identification. We propose a method for integrating features with their estimated likelihood. Matching may fail if there are condition changes such as resolution, lighting and viewpoint change between cameras. Different features are affected by different conditions. Thus it is important to choose effective features for each pair of cameras. It can be achieved by the best weights of the featuresfor the person re-identification with positive and negative training data. However in this approach, when temporal condition changes occures within a camera, we have to prepare training data under the condition. It is very difficult to prepare such training data beforehead, especially positives for all cameras. We use the distribution of the distance between features of two observed persons instead of the positive and negative training data. Using the distribution, we estimate the likelihood of a pair of observed persons being same. To estimate the likelihoods, we have to know the probability of whether they are the same person or not when a distance of a pair of persons is given. We get the probabilities and the likelihoods by estimating the two distributions of distances of positives and negatives. Experimental results show the effectiveness of the proposed method; gaining performances of matching people.