Représenter pour suivre : Exploitation de représentations parcimonieuses pour le suivi multi-objets

Loïc Fagot-Bouquet 1
1 LAAS-RAP - Équipe Robotique, Action et Perception
LAAS - Laboratoire d'analyse et d'architecture des systèmes [Toulouse]
Abstract : Visual object tracking is a subject of significant relevance in Computer Vision and its practical applications are numerous and exploited in various areas. For example, it is used in videosurveillance domain or by self-driving car technologies that require a full understanding of the vehicle surroundings. Multiple Object Tracking based on the tracking-by-detection paradigm has widely benefited from the recent developments in object detection. However, object detectors sometimes give erroneous responses, like missed detections, false positives, or imprecise detections. Maintaining target identities and handling occlusions are some other issues more specific to Multiple Object Tracking, which remains a challenging problem. Many recent approaches have exploited complex appearance models to distinguish more efficiently the targets and gain in robustness. In this thesis, we have followed the same idea by considering appearance models based on sparse representations that have been widely used in Single Object Tracking. We focus on people tracking since most practical applications are dealing with this object category. The first contribution of this thesis consists in designing an online, meaning frame by frame, tracking approach that takes advantage of collaborative sparse representations to define the affinity values between the estimated trajectories and the last detections. Furthermore, different possible descriptions of the targets, either holistic or local ones, have been considered. Contrary to offline approaches that consider several frames, online approaches are not able to correct possible association errors like identity switches or track fragmentations. Therefore, we proposed for our second contribution to develop a tracking system with a sliding window, based on a MCMCDA approach, able to correct association errors by exploiting sparse representations well-suited for this specific framework. Since the dictionaries used are composed solely of detections, the quality of the representations based on these dictionaries is highly dependent on the performance of the object detector. In order to rely less on the detector quality, we consider for the last contribution of this thesis to use dense dictionaries that are taking into account all possible locations of a target inside each frame. Many quantitative evaluations were performed using usual and public datasets, notably those of the MOTChallenge, in order to provide a consistent comparison with other recent approaches. These evaluations show the gain in performances of our proposed contributions and demonstrate the relevance of the choices that had been made.
Document type :
Theses
Robotique [cs.RO]. Université Toulouse III Paul Sabatier (UT3 Paul Sabatier), 2017. Français
Liste complète des métadonnées

Cited literature [157 references]  Display  Hide  Download

https://hal.laas.fr/tel-01516921
Contributor : Christine Fourcade <>
Submitted on : Tuesday, May 2, 2017 - 2:14:25 PM
Last modification on : Thursday, January 11, 2018 - 6:26:29 AM
Document(s) archivé(s) le : Thursday, August 3, 2017 - 1:01:16 PM

Identifiers

  • HAL Id : tel-01516921, version 1

Citation

Loïc Fagot-Bouquet. Représenter pour suivre : Exploitation de représentations parcimonieuses pour le suivi multi-objets. Robotique [cs.RO]. Université Toulouse III Paul Sabatier (UT3 Paul Sabatier), 2017. Français. 〈tel-01516921〉

Share

Metrics

Record views

122

Files downloads

221