Dystonia is a movement disorder that causes unusual movements and involuntary muscle contractions affecting some parts of the whole body. Selecting drugs and doses is a highly personalized process for dystonia, requiring frequent visits to the clinic, pointing toward the need for more systematic and objective methods of collecting patient data. A deep learning-based pose estimation algorithm can be a good candidate for aiding independent clinical assessment of dystonia as it has outperformed the classical approach to human pose estimation. The deep learning-based model can help patients and physicians assess the first symptoms of neurological diseases and build low-cost solutions not only for dystonia score prediction but also to monitor the progress of the disease. Pose estimation algorithms with convolution networks have already been shown to extract relevant information about the motor signals of Parkinson’s disease from video assessments, and the calculated score correlates well with the clinical score. OpenPose algorithm was used for human pose estimation in videos of dystonia patients being clinically assessed to annotate body key points in the videos. This project explored the basic pipeline steps required to process the clinical videos, including spatiotemporal key points normalization. CNN successfully predicted neck dystonia scores to around the scores obtained from standard clinical assessment, leaving space for further validations and research with more data and methods.

Soccer dominates the global sports market, and viewers’ interest in watching videos of soccer matches is ramping up. Globally, there is a huge and constantly increasing amount of soccer game content being generated, including video footage, audio commentary, text metadata, goal and player statistics, scores, and rankings. As a large percentage of audiences prefer to follow only the major highlights of a game, the creation of multimodal (video/audio/text) summaries is of great interest to broadcasters and fans alike. In this regard, it’s crucial to provide game summaries and highlights of the major game moments. However, creating summaries and annotating events most often necessitates the use of expensive equipment and a significant amount of time-consuming manual labor. Recent advancements in Artificial Intelligence (AI) technology have demonstrated great promise in this context. The purpose of this thesis is to use AI to support an automated pipeline for summarizing soccer matches. With Natural Language Processing (NLP) tools and heuristics, the emphasis is on creating comprehensive game summaries in textual form with variable length constraints, based on raw game multimedia (e.g., video and audio streams) and, where appropriate, easily accessible game meta-data. A longformer model has been fine-tuned to output a game summary for a given textual input of game captions. This work also explores the use of game audio in prioritizing game events from a summarization perspective. In particular, the Root Mean Square (RMS) audio intensity score has been extracted and used to extract the event priority to be included in the summary.

This project, for the first time in its nature, introduces a new research paradigm of remote motion sensing for health monitoring of civil construction in the public safety domain in Nepal. Preliminary data from a piloting study from BRB encourages us to move forward with an ageing analysis of such civil structures. Students from DoECE at IOE, Pulchowk Campus will collaborate to test network configurations and hardware types. One such configuration will be a Master/Slave configuration over Raspberry Pi Server with Bluetooth 5.0 and multiple Arduino BLE Kits. The data collected will create an opportunity to study vehicle mobility and its impact on the bridge as well as over other multiple domains which we are very open for collaborations.