Department of Electrical, Computer, and Biomedical Engineering, Ryerson University, Toronto

Accurate real time positioning is the key to enable location-based services (LBS). Although the global positioning system (GPS) is widely used for localization in outdoors, the GPS usability is not satisfactory in the confined indoor environments. Unlike outdoor, indoor environments are very complex with varying shapes, sizes with the presence/absence of stationary and moving objects (e.g. furniture and people). These factors significantly alter both line-of sight (LOS) and non-line of sight (NLOS) radio signal propagation causing unpredictable attenuation, scattering, shadowing and blind spots that significantly degrade the accuracy of indoor positioning.

However, due to the high demand for LBS, significant attention has been made on the development of indoor positioning systems (IPS) recently. Typical ranging techniques based on received-signal-strength-indicator (RSSI), time-of-arrival (ToA), time-difference-of-arrival (TDoA), angle-of-arrival (AoA), and channel-state-information (CSI) have been proposed. Most ranging techniques require at least three known anchor nodes to calculate the location of the unknown target.

All these approaches suffer from multitude of challenges including poor accuracy, high computational complexity, and unreliability while, most positioning devices lack strong processing power. In addition, the ability to maintain big databases (for large scale IPS) while ensuring security and privacy, and supporting device heterogeneity at a reasonable cost are some other challenges in indoor localization.

In recent years, artificial intelligence (AI) and machine learning (ML) algorithms find good success in indoor localization. The main advantage of AI/ML approaches is their ability to make decisions effectively using observed data without accurate mathematical formulation. ML has also proven as an effective way to fuse multi-dimensional data collected from multiple positioning sensors, technologies and methods. Both supervised and unsupervised learning can be applied for fusion weight generation. However, unsupervised ML fusion technique is superior since it calculates the weights in real-time without offline training.

In localization, classifier algorithms are mainly used to extract core features of the signals. In fingerprint-method clustering is performed based on these extracted features. Feature extraction is also important for NLOS identification and mitigation. K-NN, Support-Vector Machine (SVM), Random Forest, Decision Tree and, Artificial Neural Networks (ANN) are widely used classification algorithms.

However, in complex environment scenarios where features extraction is difficult and data has high dimensionality, DL is very promising to improve localization accuracy. DL is well known for its distributed computing capability and analyzing of a huge volume of unlabeled and un-categorized data. The biggest advantage of DL algorithms is their ability to extract features from data directly without manual feature extraction. This eliminates the need of domain expertise and extraction of hard core features. Feature extraction and classification are carried out by a DL algorithm known as Convolutional Neural Network (CNN).

Many of the indoor positioning approaches are vulnerable to global positioning error and kidnapped-robot problems. The global localization problem occurs when the initial position of the target is unknown to the IPS during initialization. While kidnapped-robot problem occurs when a well-located target moves to an unknown environment. In such a challenging situation, RL proves to be the best technique to   use. As RL enables the agent to achieve a long-term objective by interacting with the environment (based on the reward and penalty process), and are able to solve problems caused by radio signal instability. Therefore, RL techniques are able to construct the map and optimize its action continuously, thus proved as promising solution for indoor positioning.