Skip to Main content Skip to Navigation

Pedestrian detection and re-identification using interest points between non overlapping cameras

Abstract : With the development of video-protection, the number of cameras deployed is increasing rapidly. To effectively exploit these videos, it is essential to develop tools that automate monitoring, or at least part of their analysis. One of the difficulties, and poorly resolved problems in this area, is the tracking of people in a large space (metro, shopping center, airport, etc.) covered by a network of non-overlapping cameras. In this thesis, we propose and experiment a new method for the re-identification of pedestrians between disjoint cameras. Our technique is based on the detection and accumulation (during tracking within one camera) of interest points characterized by a local descriptor. We present and evaluate a keypoints-based method for modeling a scene background and detecting new (moving) objects in it. Then we present and evaluate our method for identifying a person by matching the interest points found in several images. One of the originalities of our method is to accumulate interest points on sufficiently time-spaced images during person tracking, in order to capture appearance variability. We produce quantitative results on the performance of such a system to allow an objective comparison with other features (SIFT, Color, HOG). Finally, we propose and test possible improvements, particularly for the automatic selection of moments or interest points, to obtain a set of points for each individual which are the most varied and more discriminating to those of other people. This probabilistic variant of our method brings tremendous improvement to performance, which rises at 95% first rank correct identification among 40 persons, which is above state-of-the-art.
Complete list of metadata

Cited literature [142 references]  Display  Hide  Download
Contributor : Bibliothèque MINES ParisTech Connect in order to contact the contributor
Submitted on : Wednesday, February 16, 2011 - 10:43:58 AM
Last modification on : Wednesday, November 17, 2021 - 12:30:58 PM
Long-term archiving on: : Tuesday, May 17, 2011 - 2:50:11 AM


  • HAL Id : pastel-00566417, version 1


Omar Hamdoun. Pedestrian detection and re-identification using interest points between non overlapping cameras. Computer Vision and Pattern Recognition [cs.CV]. École Nationale Supérieure des Mines de Paris, 2010. English. ⟨NNT : 2010ENMP0055⟩. ⟨pastel-00566417⟩



Record views


Files downloads