Renforcement de la localisation piétonne par détection visuelle via des informations géométriques et spatiales contenues dans une carte

This paper presents a technique to reinforce a proprioceptive system of indoor pedestrian location, assisted by a camera,embedded in a phone. A YOLOv5 neural network is trained to detect landmarks present in the images. The main contributionof this paper is to allow the system to identify false positives by introducing geometric and spatial constraints drawn from thebuilding map. This is accomplished by a Particle Filter with management of particles in groups validating or not the detection in amulti-hypothesis mode.