Deep Learning Framework for Target Localization and Tracking in Error-prone Environment

Abstract

The use of Internet of Things (IoT) in environment monitoring has led to the development of Smart Environmental Monitoring (SEM) paradigm. Target Localization and Tracking (L&T), that determines the underlying cause of environmental occurrences, is an important aspect of SEM. To prevent an environmental event in becoming a potential disaster, swift and early localization of its source and its active tracking, becomes crucial. L&T is performed with the help of IoT sensors, which are typically deployed in hazardous environments and are not easily serviceable, making energy efficiency a requirement for SEM systems. Furthermore, IoT sensors may provide anomalous readings due to low battery, low sensor sensitivity, faults, or malicious attacks. This alters the decision-making process and leads to inaccurate localization of the source. Hence, detecting and dealing with anomalies, while utilizing a limited number of sensor nodes, are key factors in addressing reliability and energy efficiency issues.

The current localization works- (1) are iterative in nature, and employ redundant nodes, which causes slow localization and leads to unnecessary energy consumption, (2) do not efficiently address anomalous readings, thus incorporating substantial localization errors, (3) are mostly application specific, which cannot be applied to other sensing tasks. Therefore, this work proposes a comprehensive deep learning-based system that- (1) repopulates missing data from inactive sensors, (2) detects and rectifies anomalous data, and (3) instantly localizes a target. The efficacy of the proposed approach is validated in a radioactive environment and compared to existing benchmark. The results show that the proposed approach achieves rapid, precise, and efficient localization.

The current tracking works- (1) mostly perform continuous localization, which causes large overhead on the system, (2) only predict one future location, and (3) do not perform continuous tracking. Therefore, this work proposes an energy-efficient continuous predictive tracking framework, which consists of two phases: 1) trajectory prediction using transformers network for predicting multiple future locations at once, and 2) path correction using polynomial regression for enabling accurate continuous tracking. The results show that the tracking system consumes very less energy while delivering high tracking accuracy in comparison to other deep learning models.

Date of Award	Dec 2022
Original language	American English
Supervisor	Shakti Singh (Supervisor)