Mobile crowdsensing applications for intelligent parking and mobility

Due to the worldwide rapid urbanization and rising private vehicular mobility, most cities face problems such as traffic congestion, parking deficit, and cruising for parking, leading to resource wastage and environmental costs. This includes the first/last-furlong problem for personal vehicular mobility in public places i.e. the challenge to find/locate a parking space near the destination. Intelligent mobile sensing, which fuses the computing/sensing capabilities of mobile devices with machine learning, holds a great potential towards solving such urban mobility challenges. Opportunistic Mobile Crowdsensing (MCS) has also emerged as a new paradigm that enables intelligent mobility applications and services. Mobile computing research has addressed the on-street/outdoor traffic and parking problems to a great extent. However, the occupancy estimation of indoor parking garages remains mostly unsolved - especially in garages that lack vehicle-tracking sensor infrastructure and/or real-time information delivery. In fact, the major bottleneck that precludes us from reusing successful outdoor schemes into parking garages is the inadequate indoor localization potential, due to unavailability of infrastructural support such as GPS/WiFi especially when underground. In addition, enterprises using MCS face concerns regarding the scalability of their proven system deployments to newer venues; this typically calls for extra data collection for training/tuning, which can be prohibitively expensive to reach a certain spatial urban coverage required to bootstrap a viable service. Moreover, drivers who park their vehicles in such garages often face an embarrassing problem of forgetting where they parked. Most existing indoor solutions rely on infrastructure such as WiFi or BLE, whereas current smartphone-only proposals mostly require major data collection and training efforts per garage, making them impractical to adopt at indoor parking garages. In this thesis, we address these issues by proposing crowdsensing methods, implementing system prototypes and endorsing them through empirical evaluation. In the first part of this thesis, we focus on the parking occupancy inference problem for indoor parking garages without precise localization. As we cannot use mobile crowdsensing for direct occupancy counting as in the existing proposals for outdoor parking, we propose ParkGauge+ as a scalable mobile crowdsensing system to indirectly infer the occupancy of parking garages from temporal parking characteristics reported by commodity smartphones, instead of counting the parked vehicles. ParkGauge+ adopts mostly low-power sensors (accelerometer, barometer, gyroscope) to detect driving states and driving contexts and infer parking characteristics, which are utilized to infer the occupancy using Supervised Learning (SL). To achieve a scalable performance and improved spatial coverage of urban deployments, we further propose to employ Transfer Learning (TL) methods to transfer the parking occupancy inference concepts from existing ParkGauge+ deployments. This minimizes any extra data collection efforts and costs for new installations. Through extensive experiments at commercial parking garages, we evaluate and demonstrate the potential of our methods that can achieve normalized error (NRMSE) of 0.09 and 0.15 respectively for SL and TL. In the second part of this thesis, we directly address the indoor vehicular localization problem in GPS/WiFi-deprived environments such as indoor parking garages and propose a novel light-weight smartphone-only solution called ParkLoc. ParkLoc exploits the inherent planar graph structure of the navigable paths in parking facilities, to match a vehicle trajectory onto a sub-section of the map, by modeling them as sparse directed graphs. By posing the indoor parking localization task as a subgraph isomorphism problem and exploiting an approximate graph matching method, ParkLoc can track a vehicle in real-time with a median error of 4.8m and localize a parked vehicle with a median error of 2m from the nearest parking space. Furthermore, ParkLoc adopts the GraphSLAM algorithm from robotics research and applies it into the mobile sensing environment; it learns the map graph from the observed trajectory graphs and a given set of bootstrap (seed) landmark nodes, in a semi-supervised manner. A key benefit of our approach is that ParkLoc works off-the-shelf without any expensive on-site training or sensor data collection per garage. A comprehensive evaluation of ParkLoc through extensive experiments performed in 4 different garages reveals the promising performance of our graph-based approach for both localization and mapping. Doctor of Philosophy