HKUST |
Development of BIM-assisted Access Point Placement Optimization and Deep Learning based Multi-floor Identification Algorithms for Enhancing Indoor Positioning to Support Construction Applications |
Thesis |
08/2019 |
Kenneth Chun Ting LI |
Over the past decades, indoor positioning has been drawing wide attention in different fields of engineering. Indoor positioning technologies are complementary to the mature outdoor positioning technology such that the indoor positioning technologies can provide a real-time positioning service in any environment where there is a blockage of GNSS signals. In the fields of construction and facility management, indoor positioning technologies enable promising applications that can considerably enhance the productivity, efficiency and safety on construction sites, supporting five major applications, which are (1) construction safety management, (2) construction process monitoring and control, (3) inspection of construction structures and materials, (4) construction automation with robotics, and (5) the use of building information modelling (BIM) technology for construction progress management.
Currently, there is no single perfect indoor positioning system that can perform optimally under any circumstances. In addition, due to the large variety of indoor positioning technologies and principles, as well as the complex and dynamic environment on construction sites, developing suitable indoor positioning systems on construction sites is a challenging task. Applying indoor positioning systems is essentially user-oriented and environment-specific. This thesis thus analyses the challenges to apply indoor positioning systems on construction sites, and then proposes six indoor positioning performance metrics, namely APP-CAT, for evaluating suitable on-site indoor positioning systems. Subsequently, the top 10 indoor positioning technologies, which are selected according to their evaluation results using APP-CAT and their popularity amongst the indoor positioning literature studies, are thoroughly discussed and compared. The promising recent trends of developing on-site indoor positioning systems, such as infrastructure-free positioning, collaborative positioning, game theory positioning, and device-free positioning, as well as integration of indoor positioning technologies with BIM models, are also highlighted. In this research work, the comprehensive discussion of current development in indoor positioning from different aspects is intended to help academics, researchers, and industry practitioners develop high-performing and suitable on-site indoor positioning systems for supporting various engineering and construction applications.
Among various positioning technologies, Wi-Fi fingerprinting has emerged as a popular technique due to the wide coverage of Wi-Fi signals and its high compatibility with smartphones. Wi-Fi fingerprinting utilizes the patterns of the Wi-Fi signal strengths, which are measured by the Received Signal Strength Index (RSSI), for position estimation. Normally, Wi-Fi access points are placed arbitrarily, which causes a poor positioning accuracy. In fact, positioning accuracy can be considerably enhanced by optimizing the access point (AP) placement strategy. In light of the high popularity of Wi-Fi fingerprinting and the liberty to design AP placemnent strategies on construction sites, this thesis aims to conduct AP placement optimization is by finding the optimal AP placement strategy that maximizes the distinctiveness between individual Wi-Fi fingerprints in a 3D virtual environment. The use of BIM technology provides 3D geometric and semantic information to accurately reproduce the virtual environment for realistic simulation of Wi-Fi signal propagation. Wi-Fi signal propagation is usually modelled by a modified indoor radio wave path loss model, but such models cannot easily consider the multipath effect in an indoor environment. Therefore, in this thesis, an accurate Deep Belief Network (DBN) based path loss model, which considers the multipath effect emulated by the ray-tracing method using particle swarm optimization (PSO), is proposed and implemented to predict the indoor Wi-Fi signal strengths. Based on the results of the simulation, the optimal AP placement strategy as well as the geometrically-constrained optimal AP placement strategy can be obtained by using the genetic algorithm (GA). The test results in a university library have shown that the developed AP placement optimization algorithm could consistently enhance the accuracy of 3D indoor positioning under the circumstances of different numbers of APs and the presence of geometric constraints.
Facility management is often performed in a multi-floor indoor environment such as shopping malls and airports. However, one of the major challenges facing the received signal strength indicator (RSSI) based fingerprinting is the inability to perform accurate indoor positioning in a multi-floor environment, despite their popularity. The multi-floor environment poses a large challenge to RSSI fingerprint-based indoor positioning because the uniqueness of RSSI fingerprints is largely lost in a multi-floor environment, especially when ring structure exists in the building. Such a ring structure is commonly found in large airports and shopping malls. In this thesis, in light of the analogy between visual images and a radio map, a novel twofold multi-floor localization algorithm based on convolutional neural network (CNN) is developed to perform robust and accurate multi-floor localization. To support the twofold CNN model and to improve the localization accuracy, the similar selective search algorithm and data augmentation algorithm are proposed. Lastly, with the support of inertial measuring units (IMUs), the snapping algorithm is proposed to convert a random trajectory to a grid shape for the purpose of localization. Per the validation results, the proposed multi-floor localization algorithm is capable of identifying on which floor the user is located such that the “floor jumping” problem is mitigated, and thus the overall indoor positioning accuracy on RSSI fingerprint-based indoor positioning is substantially improved during indoor navigation.
To summarise, this thesis provides a comprehsive review of the top 10 indoor positioning technologies for their usage on construction sites, and aims to develop a BIM-assisted access point placement optimization and deep learning based multi-floor identification algorithms for enhancing indoor positioning to support construction applications for both construction management and facility managment. |
N.A. |