Individual capstone project, January to May 2026Advisor: Prof. Meng-Hsueh Li
My earlier fall detector ran on a Jetson but its latency was erratic, and I could not say which stage was slow. In this project I designed a small 1D-CNN that fits on an FPGA, wrote a convolution accelerator in Vitis HLS, and measured the same model on the same 92-sample test set on a PC, a Jetson Nano, and a PYNQ-Z2. Accuracy was 95.65% on the PC and Jetson and 91.30% on the FPGA. Pose extraction took 96.7% of end-to-end time, so shrinking the classifier further would not help much. The FPGA did not make the pipeline faster overall, but its latency jitter was 0.025 ms against 0.278 ms on the Jetson.
Individual capstone project, January to June 2025Advisor: Prof. Meng-Hsueh Li
A camera loses the person in poor light or behind furniture, and a microphone is confused by a television. DST-FallNet runs both on a Jetson Nano with 2 GB of memory, entirely offline: pose keypoints from MoveNet and an LSTM on audio features each give a fall score, the two are combined with Dempster–Shafer theory, and a posture classifier is called only when the fused score crosses a threshold. Fusion itself takes under 0.1 ms. Pose estimation was the slow step at three to four frames per second, which led to the co-design project above.
Part of the VOICE project, May 2026 to presentWith Dr. Yu-Tai Lo and Prof. Wen-Hsiang Lu
A Flutter prototype for older adults that presents advance care planning content by readiness stage, with Taiwanese read-aloud, large type, and two-choice questions to keep it easy to operate. It is positioned as a tool for starting the conversation and deliberately leaves out online signing.
Team of four, September to October 2025Advisor: Prof. Li-Fang Chou
First place in the Smart Cultural Tourism track of the Visioning the South 2025 proposal competition. The proposal turns a walk through old Tainan into a city game with tasks that lead visitors into side streets and long-standing shops. I was the lead proposer and handled the concept, the app's features and flow, the slides, and the final pitch.
Touchless medical image viewer for sterile settings
Course project, team of twoIntelligent medical information systems, NCKU CSIE
During surgery, touching a mouse breaks sterility. This system lets a surgeon browse images by hand gesture, voice, and proximity using a Raspberry Pi and a PC. Hand landmarks from MediaPipe are mapped to four gestures, and a gesture is accepted only when four of five consecutive frames agree. Voice commands go through a Mandarin–Taiwanese speech recognition service, and an ultrasonic sensor enables control only when someone is within a meter. I wrote the programs on both machines and did the wiring and integration.
Adaptive ground calibration cane
Course project, team of fourMicrocontroller system design, NCKU EE
A conventional cane finds a step only when it touches the edge. This one measures the distance to the ground ahead with an ultrasonic sensor and warns by vibration and sound when the distance suddenly grows. One button press on level ground sets the baseline for the user's height and grip, and an inertial sensor corrects for the cane's swing. I wrote the STM32 firmware and the ESP32 dashboard, and did the circuit and assembly.
Workshop project, team of four from three departmentsE Ink e-paper design workshop, 2025
Paper triage tags cannot be updated when a patient's condition changes, and they do not survive water or blood. The wristband uses e-paper, which keeps its image without power and draws current only when it is refreshed. A responder picks the triage level and enters treatment notes on a phone, and the band's display follows. I built the hardware prototype and firmware and the phone interface.
