Deep neural networks (DNNs) are becoming the core components of many applications running on edge devices,especially for image-based analysis, e.g., identifying objects, faces, and genders. While very successful in resource rich environments like the cloud of powerful computers, utilizing Deep Learning on edge devices for inference is not used to great extend. The resource constraints of edge devices present bottlenecks that prevent all but small networks from on-device inference. Existing solutions like compression or off-loading to the cloud sacrifice either model accuracy or inference latency. To counter the aforementioned shortcomings, Masa and EdgeCaffe are proposed as a solution. The process of DNN inference by existing frameworks is to loaded and execute a model in its entirety. As this ignores the resource limitations that are inherent to edge devices, a new framework EdgeCaffe is proposed. EdgeCaffe is able to execute models partially and thereby reducing the memory footprint of the model. Additional, EdgeCaffe allows for coordination between models when a multi-model DNN inference is executed. Masa , a memory-aware multi-DNN scheduling algorithm, featuring on modeling inter-and intra-network dependency and leveraging complimentary memory usage of each layer. Masa can consistently ensure the average response time when deterministically and stochastically executing multiple DNN-based image analysis. With extensive evaluations, Masa can consistently ensure the average response time when deterministically and stochastically executing multiple DNN-based image analyses. We extensively evaluate Masa on three configurations of Raspberry Pi and a large set of popular DNN models triggered by different arrival patterns of images. Our evaluation results show that Masa can achieve lower average response times by up to 8× and 16× for deterministic and stochastic arrivals respectively on devices with small memory, i.e., 512 MB to 1 GB, compared to the state of the art multi-DNN scheduling solutions.

Most wireless devices use the radio frequency spectrum which is reaching its limits, as almost all frequencies are in use. Devices which communicate on the same radio frequencies are interfering with each other and this creates problems in the communication. Currently, there are around 10 billion of these devices. This number is rapidly growing and this increases the issue of interference. One of the possible solutions for future wireless communication is a communication technology called Light Fidelity (LiFi). This technology uses light waves to communicate rather than radio waves. Since light operates outside the common radio frequencies, LiFi does not introduce additional interference to today’s commonly used radio communication technologies. Signify is the first company with a commercially available LiFi system and has branded their system Trulifi. The Trulifi network can consist of multiple Access Point (AP)s and End Point (EP)s. These devices have a cone­like coverage area in which devices can communicate with each other. Outside this coverage area, communication is not possible. As with other communication technologies, devices within a LiFi system are also affected by interference. In the overlapping coverage areas between two APs two types of interference can occur. Downlink interference occurs whenever an EP receives messages from multiple APs at the same time, while uplink interference occurs whenever an AP is able to receive messages from an EP which is connected to a neighbouring AP. To overcome this issue Signify uses a so called LiFi controller (LC). The main task of the LC is to create an interference free communication schedule for all APs in the LiFi network. This schedule is based on the interference reports received from the APs and EPs. Validating large scale LiFi systems can become complicated, in terms of high costs involved and in the very large spaces needed to install such a large scale system. To ensure the LC operates as specified, a stress­test simulation model was designed by Signify. The stress­test simulation’s sole purpose is to ensure the LC keeps operating as specified in the worst case situation. E.g. all the possible APs, all the possible EPs and all the interference that can occur. This test is, however, not realistic as these worst case situations are not common in real­world deployments of the system. In this thesis, the problem of creating a realistic Trulifi simulation model is addressed. A literature study has been performed to investigate if existing models can be used. The conclusion of this literature study shows that a radio model needed to be designed which addressed the needs of the Trulifi system. This model has been designed with a high level of abstraction, while keeping the simulation as realistic as possible. To validate the new radio model, the Trulifi simulation has been compared to a real­world Trulifi system. Once this radio model was validated, the scheduling was compared based on different configuration sizes, as the main goal of the thesis is to create a realistic simulation model with which the Trulifi system can be validated. For this validation, the Trulifi simulation was compared to both the real­world Trulifi system and the stress­test simulation.

A start-up creates videos which users can watch to experience their running or cycling activity all over again. Currently, the company depends on external data sources to generate a video. To be less dependent on these sources the company wants to create their own tracking solution. This solution has to fit in their existing smartphone application available for iOS and Android. The company wants to remain flexible, therefore the tracking application has to be developed in such a way that it can also be used in other products the company might develop in the future. As a goal, the data has to result in visually pleasing videos for a large user base. Based on an experimental app developed during the research phase, raw smartphone GPS data was found to be unsuitable for video rendering. To improve this data, a Kalman Filter is used, in combination with a smoothing algorithm. The system has been designed to allow code sharing between iOS and Android where possible. The system has been implemented in Objective-C, Java, and TypeScript. Separating the system in three blocks enables code reuse which improves maintainability of the system. The filter has been integrated as shared code in the TypeScript implementation, which allows filtering to happen on the device. The user of the React Native Module developed has freedom to retrieve the unprocessed and processed data. The system has been tested by means of unit tests in all three programming languages used. Tests have been executed using a continuous integration server, testing each pull request against the current code base to ensure quality. Part of the testing phase includes the React Native Module to be implemented in the client's smartphone application to demonstrate its use. The application has been sent to a number of test participants to collect data from different routes and activities. The project can be seen as a success since all important requirements have been successfully implemented.

The e-sport of drone-racing involves human pilots to race against time. Recently, drone races have also gone fully-autonomous. As a result, these agile robotic platforms not only pose challenges of flying fast to the participating pilots but also create challenges for the flight control computers. As a result, the concept of autonomous drone racing has gained significant attention from research groups around the world. These races aim to push the boundaries of perception and control algorithms, while simultaneously mitigating the real-world uncertainty of execution on autonomous systems. While perception algorithms face challenges due to limited feature detection, high motion blur and computational requirements, control algorithms face challenges of convergence to the desired trajectories that are planned out in the race arena. This thesis addresses the challenge of control for racing, which is responsible for guiding the drone to design and track desired trajectories for fast flights. The control sub-modules of racing drones are responsible for generating trajectories for fastest possible flights and also for obeying these generated commands. Additionally, the requirement of limited algorithm complexity is added to match the philosophy of computationally efficient algorithms at the Micro Air Vehicle Laboratory. However, to address the requirements of these control sub-modules, the prerequisite of accurate state estimation always persists. Assigning control actions to a robot without information on the current state of the robot is rather unwise. As a result, this thesis first aims to perform accurate state estimation before designing controllers for time-optimal trajectory tracking. Again, another constraint of using only a single sensor (i.e. the Inertial Measurement Unit) is added to make the drone race in GPS denied environments. As a result, the goal of the thesis is two-fold i.e. making accurate state estimators while using limited sensors and designing optimal controllers for taking the quickest trajectory through the arena. To achieve the goal of accurate state estimation, existing techniques are studied. Several features from each of these methods are selected to design a new estimator. To achieve the goal of time-optimal trajectory generation, firstly, the flaws of traditional control methods are pointed out. A new optimal-control technique is proposed, which makes use of fundamental principles dating back several decades. This principle is then fused along with present-day optimization solvers. Finally, the proposed state estimation and control algorithm are compared against prior (benchmarked) techniques in the area. Compared to existing optimal control techniques, the proposed algorithm leads to faster trajectories and consumes less computational power onboard.

Due to the current trend of urbanization more people will come and live in cities than ever before. This increases the need to monitor the urban environment, to be able to improve the living conditions of the urban dwellers. Our project combines the need to monitor the local environment, with a portable, self-powered and wireless weather station. The final design consists of a solar powered and WiFi-connected weather station with autonomous functionality for one year. For our research, the focus was on the Netherlands. Part of such a project is the power management which will be dealt with in this thesis. In the thesis, different maximum power point tracking algorithms will be discussed and compared. Furthermore, some more research is done in the hardware design, the use of different evaluation boards and battery configurations. Finally, a prototype using the incremental conductance algorithm is constructed and tested. Simulations lead to an efficiency of around 80%, which means autonomous functionality for a minimum of one year. The physical system had a lower efficiency, but autonomous functionality for one year was still achieved.

The off-side rule is the most complicated rule in soccer. Today, off-sides are determined with the help of referees running on the sides of the field or with the help of cameras. However, for lower league games the installation of a camera based system is very expensive. This work proposes a sensor based system that can be worn by players which can automatically detect off-sides. In order to create such a system, two aspects had to be researched (a) localisation and (b) ball detection. This work provides an evaluation of an ultra-wide band system in outdoor conditions to solve the localisation problem and various proximity based sensors to solve the ball detection problem. The primary research objective of this work was to find the pit-falls of the ultra-wide band system with respect to our application. It was found that the tri-lateration algorithm implemented on the devices contributed to a lot of errors in localisation accuracy and the overall system suffered from significant packet loss. A custom tri-lateration algorithm along with hardware improvements are proposed. Our results show that the localisation errors are reduced by 56 % and improved system suffers from nearly no packet loss. As for ball detection, results show that ToF sensors provided the best accuracy in detecting the ball.

For the development of automatic People Counting systems, radar is increasingly becoming a popular technology because of the increasingly stringent privacy requirements for people demographic information and the requirement to operate in a challenging environment. Because of the complexity of multi-target movement and the diversity of application scenarios, Radar-based People Counting methods are required to have sufficient robustness. However, based on the review of the current literature, the grouping phenomenon (i.e., multiple individuals moving close together as a single group) was not often considered in the experimental scenarios. This thesis aims to study one of the most complex motions of individuals, grouping, and address the People Counting problem more in general, including the cases of grouping of multiple individuals. After studying the characteristics of Group People (defined as a group of people sharing neighboring, adjacent locations and moving together), with the help of multiple-input-multiple-output (MIMO) frequency-modulated continuous wave (FMCW) Radar, the combination of the Range-Azimuth map and spectrogram/cadence velocity diagram (CVD) is proposed to solve Group People Counting. Algorithm-wise, there are two categories of existing Radar-based People Counting methods, namely, tracking for counting methods and feature-based counting methods. It was found that these two categories of methods have complementary strengths. Thus, the proposed method combines the tracking for counting approach and feature-based counting approach into a new processing pipeline to estimate the number of people in each group in the scene while tracking each group. Based on it, the proposed method achieves "Beyond classification", which is the output the unlabeled classes not defined at the training stage. Moreover, compared with other state-of-the-art (SOTA) Radar-based People Counting methods, the proposed method outperforms them, and thus it is proved that the grouping problem can be solved in the Radar-based People Counting field.

Visible Light Communication is a method of wireless communication that avoids the oversaturated frequencies used by radio communication. Prior research typically uses the camera on smartphone for communication, but using a camera is energy-intensive and inefficient. Some alternatives are photodiodes and ambient light sensors, however the former is often not available on phones and the latter is too slow to be useful. This thesis introduces a system, JAB, which combines two sensors to achieve high throughput while keeping energy usage low. The impact of distance, modulation frequency and light are measured on the system implemented on an Android smartphone. Overall, the system is usable and does greatly improve the energy effiency of using VLC. An important next step is creating a custom hardware setup with a kernel driver implementation for a full system test.

Community question answering (CQA) platforms provide a social environment for users to share knowledge online. Users can submit complex and subjective questions on CQA platforms and then derive the desired answer from other community users. A large number of user-generated data has been produced by various CQA sites (e.g., Quora, StackExchange) and been used in different CQA researches. Question retrieval task is one of the popular CQA tasks aiming at solving the overloading issue of CQA platforms and increasing user satisfaction by reducing their waiting time. A question retrieval system is expected to automatically retrieve similar questions from the CQA archives regarding a new question, and the answers to similar questions are returned to users directly. Different information retrieval (IR) approaches have been proposed for question retrieval task ranging from the conventional retrieval models to the learning to rank models and neural ranking models. However, the IR community is now facing the issue of overusing the weak baselines. Thus, it is hard for researchers to identify the reported improvement of the newly-proposed methods, which greatly impedes the development of the community. Some researchers have already proposed several competitive baselines for ad-hoc retrieval task, but currently, the proposals of strong baselines for question retrieval are still not enough. Thus, this work targets on identifying the suitable baselines for question retrieval task on different datasets. We conduct an empirical comparison among different retrieval models on two representative datasets and analyze the performance of models on different question sets. Analyzing on CQA questions is challenging since the CQA questions are more diverse and complex, compared to the questions on traditional question answering (QA) (e.g., Wikipedia) system as well as the queries on traditional search engine (e.g., Google). Our work investigates the impact of the question from two perspectives. We first display how retrieval performance changes on various question sets (e.g., questions with different lengths and different levels of specificity) and then explain the reasons for the performance changes. Moreover, we conduct an error analysis to reveal the hard types of questions for different retrieval models on two datasets. In order to overcome the existing weakness of the retrieval models, we further select two techniques that have already proven effective in other retrieval tasks. We hypothesize that the two techniques can also be useful on question retrieval task. We then implement the two techniques on our datasets to validate the hypothesis. Our findings show that one of them can not help to enhance the retrieval effectiveness of models due to the different characteristics of task design while another technique successfully demonstrates the additive effectiveness gains. Based on our findings, we find out the suitable baseline models on different datasets as well as emphasize their relative strength and limitation. We believe our work can provide useful guidance on how to select an appropriate baseline for future works on question retrieval.