Abstract:
Person re-identification is one of the most critical activities in the security area, specifically in video-surveillance since it has wide applications such as access control, people tracking and behavior detection. In this paper, a system of Re Identification of people through 3 stages is proposed. The first one, detection and segmentation of people using Mask-RCNN method, the second, feature extraction with convolutional neural networks (CNN), and finally, the identification of people in different places with a multi-input neural network model and an output composed of a CNN. The model uses two types of descriptors based on soft-biometric appearance features, body silhouette and color in RGB space. These are treated and handled independently by deep learning techniques, which allows to generate as output the identification of persons. The experiments are carried out with a dataset created in a controlled environment by capturing videos with 2 counterposed cameras. Through a detailed comparison and the analysis of different models with different accuracy metrics, it can be indicated that the fusion of the silhouette and color features improve the solution robustness, than when treated individually. In terms of accuracy metrics, training time and validation, the multiple input model is the best evaluated in our experiments.
