People detection, tracking and re-identification for smart video surveillance

Resumen

This thesis addresses the problem of people multi-tracking in a non-overlapping camera network by means of different computer vision techniques. People multi-tracking plays a very important role in smart video surveillance systems, which currently are one of the most active research fields in computer vision. The collected video data can be very useful for multiple applications, covering from the extraction of person behavior in a social environment to the activation of an alarm for preventing dangerous events. However, this video data is frequently underused due to the amount of information provided by the camera network, which must be manually processes. Looking towards this problem, different contributions are introduced to make more feasible monitoring video surveillance tasks for the operator, automatically analyzing the people movement in an environment. First, we focus on the people tracking problem in a single field of view where the estimation of the person position and his/her trajectory in the image are achieved. To do this, two novel description models are proposed based on silhouette, size, motion and height, allowing the person detection over different camera configurations. Afterward, the person tracking problem is extended to the non-overlapping camera network case whose main objective is to provide a single identifier for each person. Regarding to this, a new method is proposed to carry out the re-identification process in a gradual and iterative fashion attending to the orientation distance. A pairwise feature dissimilarity space is introduced to train a support vector machine classifier (SVM). In the post ranking person re-identification field, a discriminant context information analysis module is proposed to optimize the initial ranking computed using any baseline models. Novel content and context information are defined to be used in a discriminant feature analysis in order to remove the common information from a set of feature vectors with similar appearance. Finally, we present experimental results for each method over datasets recorded in real scenarios and comparisons with state-of-the-art methods.