[time]
2020-08-27T10:36:31+02:00

[team_name]
Next-Newbie Reckoner (NNreckoner)

[team_institution]
Singtel Cognitive and Artificial Intelligence Lab for Enterprises at NTU (SCALE@NTU); Nanyang Technological University, Singapore. 

[logolink]


[system_name]


[website]


[track]
3

[reference_person]
Seanglidet Yean

[email]
seanglidet.yean@ntu.edu.sg

[description]
Brief System Description (100-200 words)

Our team has been working on the localisation and tracking in indoor and outdoor environment. It is to provide an augmented assistant to caregiver to locate the people with dementia. In the indoor environment, it has been a challenge to estimate the location with GPS. In our work, we have adopted grid-based approach to train the Random Forest models with an end-to-end pipleline in order to predict the building, floor and location grid using the WiFi Received Signal Strength. In addition, we are working on the Deep Neural Network models as well as feature selection methods to further improve the accuracy. 

In this competition, our main objective is to incorporate multi-source of information such as Bluetooth, Smartphone sensor-fusion techniques (e.g. pedestrian dead reckoning) to enhance our method. (127 words)

----------------------------------------------------------
System Description (500-2000 words) 

Nowadays, the need for indoor localization is increasing as it has many possible implementations in many sectors, e.g. navigation, health care, etc. In order to obtain an accurate indoor location, fingerprinting is the most commonly used method. The problem with fingerprinting approach is the high variation of RSSI values, resulting in erroneous location estimation. Machine learning approach is a new alternative to fingerprinting approach that aims to solve this problem. 

In our work, we adopt grid-based approach to train the Machine Learning models with an end-to-end pipleline to autotune the models hyperparameter. In addition, we explore the impact of features by studying the default RSSI values for undetected AP as well as the feature importance. It highlights the significance of selecting the data pre-processing and the AP’s RSSI combinations. RSSI data are collected from WiFi access points (APs) as WiFi access points are widely available and are at fixed location, which does not require specific hardware support. We have benchmarked our results with the publicly available dataset UJIndoorLoc (UJI) [1]. It has shown that our proposed pipeline/framework provides a resilient constructed model in localizing the grid-base location when some of the APs are not detected.

Moreover, we are working on a Deep Neural Network and transfer learning model to improve the accuracy as well as to be adaptive to a new environment. There are 3 component that are being considered: increasing volume of data for the DNN model, reduce input features to decrease complexity, and create hierarchical classification label for the hierarchical nature of building, floor and location labels. To train deep learning models, typically big data sets are required. Hence, to expand the size of the data set, data augmentation is studied; in particular, the extrapolation and aggregation technique. Subsequently, inputting the detected SSIDs to the machine learning models increase the feature space complexity which cause the data to become sparse for the model to converge. In other words, the higher dimensional feature space, the more data is required. Thus, the stacked autoencoder, as a feature extraction method, is applied. Transfer Learning is being explored in order to produce an adaptive model for the new environment as it requires fewer data, hence decrease the time consumed in the data collection process. We are looking forward to bringing the deep learning model into the test in the IPIN2020 competition dataset.

While WiFi RSSI data is commonly collected for indoor localisation since WiFi access points are widely deployed and available, the issue with WiFi RSSI data is that the fact that it is highly affected by free-space loss, reflection and multipath propagation [2]. It is crucial to incorporate the multi-source of information to increase the confident level of the prediction and improve the navigation experience. With grid-based classification model, the location update is dependent on the grid size which cause the sporadic location update. In recently, smartphone has been affordable and resourceful with the equipped with sensors that capture user’s movement. We would like to integrate pedestrian dead reckoning to complement the prediction model, especially for the movement within the grid. (510 words)

References: 
[1] J. Torres-Sospedra, R. Montoliu, A. Mart´ınez-Us´o, J. P. Avariento, T. J. Arnau, M. Benedito-Bordonau, and J. Huerta, “Ujiindoorloc: A new multi-building and multi-floor database for wlan fingerprint-based indoor localization problems,” in 2014 international conference on indoor positioning and indoor navigation (IPIN). IEEE, 2014, pp. 261–270.
[2] S. Sadowski and P. Spachos. Rssi-based indoor localization with the internet of things. IEEE
Access, 6:30149–30161, 2018.


[references]