A TinyML Deep Learning Approach for Indoor Tracking of Assets

Abstract

Positioning systems have gained paramount importance for many different productive sector; however, traditional systems such as Global Positioning System (GPS) have failed to offer accurate and scalable solutions for indoor positioning requirements. Nowadays, alternative solutions such as fingerprinting allow the recognition of the characteristic signature of a location based on RF signal acquisition. In this work, a machine learning (ML) approach has been considered in order to classify the RSSI information acquired by multiple scanning stations from TAG broadcasting messages. TinyML has been considered for this project, as it is a rapidly growing technological paradigm that aims to assist the design and implementation of ML mechanisms in resource-constrained embedded devices. Hence, this paper presents the design, implementation, and deployment of embedded devices capable of communicating and sending information to a central system that determines the location of objects in a defined environment. A neural network (deep learning) is trained and deployed on the edge, allowing the multiple external error factors that affect the accuracy of traditional position estimation algorithms to be considered. Edge Impulse is selected as the main platform for data standardization, pre-processing, model training, evaluation, and deployment. The final deployed system is capable of classifying real data from the installed TAGs, achieving a classification accuracy of 88%, which can be increased to 94% when a post-processing stage is implemented.

A Selection of Starting Points for Iterative Position Estimation Algorithms Using Feedforward Neural Networks

“…However, compared to the results of other investigations presented in the literature, such small network configurations are not surprising. For example, in [ 28 ], in a solution based on direction of arrival (DoA) estimation, the neural network consisted of two layers with a dozen neurons each, while in [ 20 ], FNN with 10 neurons in the first layer and 20 in the second one was successfully used for RSS-based position estimation. Even smaller neural networks with no more than five neurons were tested in [ 18 ] in a small positioning network based on RSS measurements.…”

“…In some of the abovementioned positioning systems with artificial intelligence that used fingerprinting principle, the results of position estimation may differ from the reference points that were used to collect the “fingerprints” because the neural network can perform some interpolation of the data. However, in [ 20 ], a neural network was used to classify measurement data into one of twenty-six possible locations in an indoor test environment. This solution used Bluetooth Low Energy (BLE) tags to transmit radio packets, which were received by five stations that performed an initial analysis of the RSS results, calculating the average value, maximum/minimum, root-mean-square, standard deviation, skewness and kurtosis.…”

Human–Robot Interaction Using Learning from Demonstrations and a Wearable Glove with Multiple Sensors

“…Notably, the During the data analysis step, the sensor data was fused before being used as input for the machine learning (ML) model. The TinyML model was developed using a preprocessing feature extraction method in the development phase [18]. Following the training phase, the ML model was tested on the test dataset.…”

Device-Free Tracking through Self-Attention Mechanism and Unscented Kalman Filter with Commodity Wi-Fi

“…Recently, significant advances in embedded systems and radio have led to the emergence of ubiquitous device-free wireless systems, which have become an important research area. Since wireless networks are ubiquitous everywhere in the range of the transmitted signal, we would be interacting with radiofrequency (RF) electromagnetic (EM) waves [ 1 , 2 , 3 , 4 ]. The device-free-based wireless systems are used in different applications such as smart homes, monitoring industrial automation, medical applications, monitoring older adults in nursing homes, monitoring prisoners in custody, and target tracking [ 5 , 6 , 7 , 8 ].…”