Collaborative navigation is the most promising technique for infrastructure-free indoor navigation for a group of pedestrians, such as rescue personnel. Infrastructure-free navigation means using a system that is able to localize itself independent of any equipment pre-installed to the building using various sensors monitoring the motion of the user. The most feasible navigation sensors are inertial sensors and a camera providing motion information when a computer vision method called visual odometry is used. Collaborative indoor navigation sets challenges to the use of computer vision; navigation environment is often poor of tracked features, other pedestrians in front of the camera interfere with motion detection, and the size and cost constraints prevent the use of best quality cameras resulting in measurement errors. We have developed an improved computer vision based collaborative navigation method addressing these challenges using a depth (RGB-D) camera, a deep learning based detector to avoid using features found from other pedestrians and for controlling the inconsistency of object depth detection, which would degrade the accuracy of the visual odometry solution if not controlled. Our analysis show that our method improves the visual odometry solution using a low-cost RGB-D camera. Finally, we show the result for computing the solution using visual odometry and inertial sensor fusion for the individual and UWB ranging for collaborative navigation.

The long-term objective of our research is to develop a method for infrastructure-free simultaneous localization and mapping (SLAM) and context recognition for tactical situational awareness. Localization will be realized by propagating motion measurements obtained using a monocular camera, a foot-mounted Inertial Measurement Unit (IMU), sonar, and a barometer. Due to the size and weight requirements set by tactical applications, Micro-Electro-Mechanical (MEMS) sensors will be used. However, MEMS sensors suffer from biases and drift errors that may substantially decrease the position accuracy. Therefore, sophisticated error modelling and implementation of integration algorithms are key for providing a viable result. Algorithms used for multi-sensor fusion have traditionally been different versions of Kalman filters. However, Kalman filters are based on the assumptions that the state propagation and measurement models are linear with additive Gaussian noise. Neither of the assumptions is correct for tactical applications, especially for dismounted soldiers, or rescue personnel. Therefore, error modelling and implementation of advanced fusion algorithms are essential for providing a viable result. Our approach is to use particle filtering (PF), which is a sophisticated option for integrating measurements emerging from pedestrian motion having non-Gaussian error characteristics. This paper discusses the statistical modelling of the measurement errors from inertial sensors and vision based heading and translation measurements to include the correct error probability density functions (pdf) in the particle filter implementation. Then, model fitting is used to verify the pdfs of the measurement errors. Based on the deduced error models of the measurements, particle filtering method is developed to fuse all this information, where the weights of each particle are computed based on the specific models derived. The performance of the developed method is tested via two experiments, one at a university’s premises and another in realistic tactical conditions. The results show significant improvement on the horizontal localization when the measurement errors are carefully modelled and their inclusion into the particle filtering implementation correctly realized.

We use motion context recognition to enhance the result of our infrastructure-free indoor navigation algorithm. Target applications are difficult navigation scenarios such as first responder, rescue, and tactical applications. Our navigation algorithm uses inertial navigation and visual navigation fusion. Random Forest classifier algorithm is taught with training data from Inertial Measurement Unit and visual navigation data to classify between walking, running and climbing. This information is used both in pedestrian navigation to do stationarity detection with adaptive threshold and in particle filter fusion to exclude visual data from during climbing. Methods are evaluated in an indoor navigation test where person wearing tactical equipment moves through a building. Results show improvement of positioning accuracy based on loop closure error on the test track especially when the movement is fast paced. The loop closure error was reduced on average 4 % in two tests when movement was slow and 14 % when movement was fast.

Ultra Wideband (UWB) signals are a promising choice for indoor positioning applications, since they are able to penetrate walls to a certain extent. Nevertheless, signal reflections and Non-Line-of-Sight (NLOS) propagation cause bias in the measured range. This ranging error can be corrected with Machine Learning (ML) methods such as Convolutional Neural Networks (CNNs). However, these ML models often generalize poorly between different environments. In this work we present an instance-based Transfer Learning (TL) approach, that enables generalizing a CNN-based ranging error mitigation approach to a new situation with only a few unlabeled training samples. The performance of the UWB error correction approach is demonstrated in a real-life infrastructure-free cooperative positioning setting.

In this paper, we study machine learning methods for recognizing the motion context of a user of an infrastructure-free navigation system. Motion context is information about whether the user is running, crawling, or lying down, for instance. This can be valuable information for the command and control of a tactical or rescue operation, and it can also be used to adapt the positioning algorithm accordingly in order to improve the positioning result. We test our approach in collaboration with the Finnish Defence Forces. With only about 5 min of training data, we are able to detect the users motion context over 93% of the time using a random forest classifier. However, our tests show that the performance of the classifier is highly dependent on the user of the system. For this reason, we experiment with different classification algorithms in order to find a user-independent classifier providing a good compromise between accuracy and computational complexity. With a naïve Bayesian classifier, we achieve 85% detection rate in case when the training data is not produced by the user. In addition, we demonstrate how the motion recognition can be used to adjust the zero velocity update threshold in order to improve the performance of a foot-mounted inertial navigation algorithm.