School of Engineering Department of Computer Science and Engineering 99 AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: SU Chen-yi / DSCT Course: UROP 1100, Fall This paper investigates the application of diffusion models for generating complete RSSI heatmaps from partial data to enhance drone navigation in search and rescue missions. We divide maps into n×m grids, compute the mean RSSI for each grid, and train a conditional diffusion model to predict full heatmaps based on sampled points. Addressing challenges in adapting MNIST-based models to custom datasets, we implemented a context-aware U-Net architecture and refined masking techniques to improve image reconstruction. Preliminary results demonstrate the model’s ability to accurately reconstruct heatmaps from limited samples, promising enhanced localization accuracy and operational efficiency for drone-assisted search operations. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: ZHANG Xingjian / COMP Course: UROP 1100, Fall UROP 2100, Spring This report presents the work that I have done during this semester’s UROP course. Our project is aimed at helping to solve the problem of human pose estimation (HPE) using mmwave and lidar data, generalizing the application of HPE with mmwave, which is currently poor in training data. Basically I did three things this semester. The first thing is to preprocess the dataset mmfi lidar. The second thing is to conduct some basic experiments on these datasets and the baselines. The third thing is to test the new methods of unsupervised learning by augmenting the data and conducting the comparison experiment. Indoor Localization and Mobile Computing Supervisor: CHAN Gary Shueng Han / CSE Student: HARTANTO KWEE Jeffrey / COMP Course: UROP 1100, Spring This report presents two projects focused on applying state-of-the-art 3D reconstruction technologies for indoor space reconstruction. The first project involves modelling large indoor spaces using crowd-sourced images by combining separate 3D reconstructions together to form a cohesive 3D model. Modern machine learning-based 3D reconstruction tools such as NetVLAD, SuperPoint and Hierarchical Localization were employed for efficient keypoint matching and pose estimation across disjoint image sets. The second project involves updating outdated 3D reconstructions by incorporating new images and integrating Visual Language Models (VLMs) for better scene understanding. The proposed pipeline achieves this by localizing new images within the existing model, identifying outdated landmarks through image alignment and comparison, and systematically removing old images that contain these features.
RkJQdWJsaXNoZXIy NDk5Njg=