Re-identification and semantic retrieval of pedestrians in video surveillance scenarios

Person re-identification consists of recognizing individuals across different sensors of a camera network. Whereas clothing appearance cues are widely used, other modalities could be exploited as additional information sources, like anthropometric measures and gait. In this work we investigate whether the re-identification accuracy of clothing appearance descriptors can be improved by fusing them with anthropometric measures extracted from depth data, using RGB-Dsensors, in unconstrained settings. We also propose a dissimilaritybased framework for building and fusing multi-modal descriptors of pedestrian images for re-identification tasks, as an alternative to the widely used score-level fusion. The experimental evaluation is carried out on two data sets including RGB-D data, one of which is a novel, publicly available data set that we acquired using Kinect sensors. In this dissertation we also consider a related task, named semantic retrieval of pedestrians in video surveillance scenarios, which consists of searching images of individuals using a textual description of clothing appearance as a query, given by a Boolean combination of predefined attributes. This can be useful in applications like forensic video analysis, where the query can be obtained froma eyewitness report. We propose a general method for implementing semantic retrieval as an extension of a given re-identification system that uses any multiple part-multiple component appearance descriptor. Additionally, we investigate on deep learning techniques to improve both the accuracy of attribute detectors and generalization capabilities. Finally, we experimentally evaluate our methods on several benchmark datasets originally built for re-identification tasks