School of Engineering Department of Computer Science and Engineering 95 Department of Computer Science and Engineering Crowd Counting with Domain Generalization Supervisor: CHAN Gary Shueng Han / CSE Student: CHEN Chenle / COSC Course: UROP 3200, Spring Crowd counting is an essential task in computer vision with real-world applications ranging from public safety to urban planning. However, counting accuracy drops significantly when models are exposed to environments that differ from their training data. This project investigates domain generalization techniques for crowd counting using MPCount, a robust model architecture. We evaluate performance across three distinct datasets—ShanghaiTech, UCF-QNRF, and DroneCrowd—each offering unique challenges such as density variation, perspective shifts, and drone-captured distortions. Our experiments show MPCount achieves competitive results in zero-shot scenarios, notably generalizing to DroneCrowd without fine-tuning. This suggests promising future directions in building crowd counting systems adaptable to real-world diversity. AI + Healthcare: Research and Development of Intelligent Systems for Medical Diagnosis and Applications Supervisor: CHAN Gary Shueng Han / CSE Student: LI Wenying / DSCT Course: UROP 1100, Summer Accurate human pose estimation (HPE) has broad applications in robotics, autonomous vehicles, etc. While millimeter-wave (mmWave) radar offers a cost-effective and privacy-friendly alternative to RGB cameras and expensive LiDAR systems, its current datasets for HPE are scarce and lack diversity, restricting its generalization in real-world scenarios. To address this limitation, we propose LEMT, a novel LiDAR-expanded mmWave training approach to enhance the performance of mmWave HPE models. Our method utilizes LiDAR HPE data, leverages the LiDAR-to-mmWave synthesis and pseudo-label estimator to expand the volume and diversity of the limited mmWave datasets. Augmented with both synthetic and pseudo-labeled mmWave data, LEMT significantly improves the generalization of all mmWave HPE models, with 14.5% and a 22.5% error reduction for in-domain and out-of-domain testing datasets, respectively. AI + Healthcare: Research and Development of Intelligent Systems for Medical Diagnosis and Applications Supervisor: CHAN Gary Shueng Han / CSE Student: LIEU Kaixuan Ryan / COMP Course: UROP 1100, Fall Our research group has been working on an innovative AI-based skipping rope counter designed specifically for primary school children. Utilizing a technology stack that includes Python, Flask, SocketIo, OpenCV, Yolo, and face_recognition, we aim to promote physical activity through engaging gameplay. This progress report will highlight the functionality of the newly implemented Yolo11 pose model, along with the recently added health report and ranking system. which provide users with valuable feedback on their activity levels and foster friendly competition. These enhancements will not only improve user experience but also encourage children to stay active while having fun.
RkJQdWJsaXNoZXIy NDk5Njg=