[time]
2020-08-25T06:21:43+02:00

[team_name]
ISSAI

[team_institution]
Institute of Smart Systems and Artificial Intelligence, Nazarbayev University

[logolink]


[system_name]


[website]
https://issai.nu.edu.kz/home/

[track]
3

[reference_person]
Huseyin Atakan Varol

[email]
ahvarol@nu.edu.kz

[description]
Short description.

We plan to utilize most of the data provided by the IPIN 2020 Track-3 competition such as  WiFi RSSI, accelerometer, magnetometer, barometer, and other sensor values. The main idea is to utilize the fingerprinting localization via the use of radio and geomagnetic pattern maps generated by WiFi RSSI and magnetometer sensor readings to train machine learning (ML) algorithms. The maps will be pre-processed in order to improve their quality. Deep learning and conventional ML techniques will be utilized in order to achieve accurate predictions. Specifically, sequence models will be investigated in great detail. To enhance predictions further, building layouts, feature fusion, ensemble learning, and pedestrian dead reckoning algorithms will be explored.


Long description.

An indoor user localization is an important component of many applications such as public safety, health care, industry, entertainment, and others. Especially, the smartphone-based localization is of high interest due to its ubiquitousness of these devices.

Over the past few decades, many effective solutions have been proposed to address the smartphone-based user localization problem. Among them, the most popular method is a fingerprinting localization. It consists of two phases: offline training and online localization phases. During these phases, usually, solutions based on ML are utilized. They map the input signal data from the built-in smartphone sensors to specific locations. The employed sensors include WiFi modules, magnetometer, accelerometer, barometer, and others.

In this competition, we plan to use most of the data provided by IPIN 2020 Track-3 such as WiFi RSSI, accelerometer, magnetometer, barometer, and other sensor values. First, we will exploit radio and geomagnetic pattern maps constructed using WiFi and magnetometer readings to train location prediction models. We will apply feature preprocessing steps to improve the quality of radio maps. The location prediction models will be based on ML methods which achieved top results on user localization tasks. We plan to implement both conventional ML (e.g. k-nearest neighbors (k-NN), support vector machines (SVM), and others) and recently proposed deep learning-based solutions (e.g. recurrent neural networks (RNN), convolutional neural networks (CNN), transformers and others). Specifically, we are interested in sequence models such as RNN and Transformers which are extremely effective at exploiting both previously received signals and predicted locations. The DL-based models will be stacked with regression modules to enable real-valued output predictions.

In addition, we plan to investigate feature fusion techniques and implement multimodal prediction models, because the combination of sensory data from different sensors might yield complementary information leading to better predictions. Furthermore, we noticed that in previous IPIN competitions the layouts of the buildings, where the data was collected, were provided. These layouts contain important information and might be beneficial for user trajectory tracking. For example, by utilizing building layout we can identify unlikely events such as two consecutive predicted positions located at different sides of a wall. Therefore, we plan to incorporate extracted information from the building layouts to our user localization system. To improve predictions further, we plan to employ ensemble learning techniques. Specifically, we will investigate different approaches for combining several ML algorithms (presumably trained on different sensory data) and identify the complementary models and sensor types. Lastly, we will fuse the aforementioned localization techniques with pedestrian dead reckoning (PDR) localization in order to enhance the short-range predictions. The PDR utilizes data from inertial measurement units (IMUs) in order to calculate the relative position of a user. According to the IPIN competition data collection method, a user will keep a smartphone in a predetermined position for most of the time. This allows us to assume that the heading offset angle is constant which means that the predictions will be accurate enough. Moreover, accumulated error due to the IMU sensors could be minimized via suggested fusion of localization methods.

Overall, the issue of accurate indoor localization will be evaluated and tackled from different viewpoints. The main focus is to experiment and combine different models in order to achieve better predictions.

[references]