Correction of pedestrian localization via map-based geometric and spatial information by visual detection of landmarks

Indoor Localization is often complemented by Computer Vision methods that provide information about the surrounding. Object detection can help a system better understand its environment and its relation to a map which refines the localization process. With the latest advances of detection models and how relatively light weight they can be, an approach to integrate these models to existing localization system can be proposed as solution to the heading drift problem while keeping hardware cost down.

This paper presents a correction method of the heading drift caused by noisy gyroscope integrated data into Pedestrian Dead-Reckoning for self-contained indoor localization solutions. This correction is done by the observations that are introduced into a Particle Filter. The observations consist into exteroceptive information extracted from remarkable landmarks present in the indoor environment. The images collected during the walk are fed through a neural network based object detector, YOLOv5, to detect extinguishers. A methodology of adaption of this network is also proposed. The handling of false positives produced by the neural network is done by applying geometrical and spatial constrains extracted from the indoor map.