New Technology Investigations, Google NestR&D of new technologies for Google Home, Google Nest or other Google productsHuman Presence detection using already-installed sensors • Build Android app for collecting sensor data• Design signal processing as preprocessing for ML• Design ML model for detecting human presence• Design Android app for detecting human presence including Tensor flow lite model Human Activity detection using weight sensors• Design and build a hardware prototype including microprocessor, load-cell sensor,• POC for basic detection using sensor data Electric power signal analyzer• Design digital filter for syncing power signal• Design algorithm for timing detection Submitted 8 potential patent idea (6 are first author)• Power controlling, solar power defect detection, earth quake detection, human presence detection, human activity detection, weight sensor product

Developing Amazon Dash Cart.Weight computation module in Easy checkout shopping cart system Own a project for weight estimation package in next generation of dash cart in designingalgorithm, create professional code, maintenance deployment. Design a new wedge detection algorithm, achieves more than 90% of accuracy fordetecting the item is wedged between lower and upper basket in the cart Design an accurate new weight measurement and calibration algorithm using multiplesensors, using combination of neural-network and digital signal processing, whichcertifies NTEP compliant. Design a new 3D position estimating method using weight sensor and IMU sensor. Proposed cloud-based dash cart system which reduce around 80% of the devicecomputation power and utilized the computational resourcesObject Recognition system in Nvidia Xavier AGX/NX Program and maintenance an object recognition system used in Amazon Dash Cart.Achieved 200% faster deployment architecture by sharing tensorRT compiled filesDNN module in DeepVision Inc, Neural Network Chip Evaluate DNN models using DeepVision SDK for next generation cart

Work as senior software engineer in headpose teamEmbedded Software program mainly for SLAM in Intel Movidius, and Nvidia Tegra including:• DNN optimization for embedded system: compressed the network size to 1/4 and increased speed to 4 times with decreasing the accuracy by less than 1%• Map management: reduced 40% of computation time with efficient use of map points memory, and reduced 40% computation time by efficient SIMD code.• DNN architecture: reduced 30% of computation time by inventing efficient distance computation network• Localization: implemented SIMD code for DNN-based localization algorithm. Tuned up multi-dimension vector search algorithm to increase speed by considering memory arrangement.• Feature description: implemented SIMD code to increase the speed of feature matching for NN-based descriptor.• Fail detection: set-up failure detection routine for input devices.• Cloud computing: building a mechanism to compute some device task in server side.• Large map merging: implemented map merge program enables to share space of multipleperson• Program maintenance: built a mechanism to share the program code in both device andserver side.

Deluca Oven (Intelligent High-speed oven) : Tensoflow, Keras, PCB design, high-current switching, Motor control. PIC, Arduino, Android, USB, Deep Learning, C++/C, Python, LinuxComputer Vision: food classification using deep learning• Design Deep CNN network in Tensorflow and Keras over Python interface.• Design next-generation oven using deep learning method, then achieve 97% accuracy todetect the food goes to oven.• Design and build state-of-the-art oven, achieve 300% faster baking time.• Design the system structure based on cost, accessibility, development speed and overalldesign.• Direct software team members to build Easy-to-understand interface in Android tablet.• Resource plan and management for completing electrical part• Design a high-current, more than 300A, solid-state switch to achieve cost down and quietswitchingToast Level Analyzer : Android, Realtime Computer Vision, Java, C++, Android NDK.Computer Vision: Fish-eye lens correction, Adaptive thresholding• Prototype for controlling heat for existing factory oven• Design toast level analyzing system to quantize the oven result.• Design Android software using internal camera to compute the toast level• Design Box, Lightning, Hardware circuit.Bun Face Detector for Industry Oven: LED drive, PhotoTransistor, Microprocessor, Embedded SoftwareComputer Vision: Line image detection, histogram intersection, decision tree• Design low-cost bun facing up/down detection system for industrial oven.• Design circuit for led, photo transistor and microprocessor.• Design PC software for visualization and machine learning

BrainAthlete (Athletes’ mental training kit) : BCI, Web Socket, Redis, MySQL.• Design a world’s first mental training kit for athlete and release it on Japanese and US market.• Communicate with Brain-Computer interface chip manufacture, Neurosky, to design and build visor type brain wave scanner.• Communicate with manufacturer to achieve mass production of the device.• Work closely with software engineer and designer to build a communication software.• Set up a real-time communication service over the net on Amazon EC2

VET (Vision Enabled Training, Intelligent surveillance for kitchen, high-level hygiene room) : Realtime process, feature tracking, feature identification, database, LinuxComputer Vision: Glove detection using color segmentation, contour feature, Washing activity using color segmentation• Design the system structure consisting of IP camera, local server, hardware sensors andinternet server.• Design glove detection algorithm using computer vision methodologies.• Distribute the information and task to camera, local server and internet server to achieve real time detection.• Set up an experimental service at high-level hygiene room in hospital at University of Michigan.• Received Kitchen Innovation Award from National Restaurant AssociationMeat Quality Report Service: realtime process, DirectX, feature tracking, C/C++, WindowsComputer Vision: Object tracking and object classification using color segmentation and contour feature, SIFT• Analyze size, fat/lean percentage of meat, inspect correctly packed on tray.• Find the best camera position and the best algorithm within limited space in the meatprocess factory.• Apply DirectX library to achieve realtime (30 frames/sec) processing.Positioning System for Factory Cleaning Robot: feature matching, database, realtime process, C/C++, OpenCV, Linux, SLAMComputer Vision: feature matching and object tracking using SIFT,• Propose a cost-effective, passive method to find the local position for cleaning robot.• Build a feature map database for a location reference.• Generate trajectory using SLAM.High Speed Computer Vision Library: embedded system, memory access optimization, C/C+ +, Linux, WindowsComputer Vision: Binarization process, pipeline process• Alternative to OpenCV library• Increase a speed by limiting the compatibility.• Small memory usage designed for embedded system• Analyze the bottle neck of speed in common computer vision methods and increase theirspeed by decreasing the memory access

MDAS (Mitsubishi Driver’s Attention Monitoring System): Embedded System, Lane detection, vehicle recognition, mass-production, FPGA, C/C++Computer Vision: Lane detection using edge, white line filter and shape matching• As a main engineer, work with hardware engineer, camera engineer and client engineer atMitsubishi Motors to design and build a world’s first drive monitoring system for truck,released by Mitsubishi Motors.• Increase the stability of lane recognition by wide range search and fast hough transform.• Increase the response time by applying energy flow method.• Increase the performance using Altera FPGA and SH-1 microprocessor.Preview Distance Control: Level-1 Self-Driving System released in 1996Embedded system, Lane recognition, vehicle recognition, video signal, Laser Radar, mass production, C/C++Computer Vision: Lane Detection, Distance computation using 2D-3D conversion• Work with a main software engineer to find a lane recognition and vehicle detectionalgorithm of a first-level autonomous driving product released by Mitsubishi Motors.• Learn to design embedded system and hardware design using H8-500 micro-processor.• Learn to design analog and digital circuit for video signal• Learn to bring a prototype product to mass production• Communicate with Laser radar team to integrate distance map from radar and white lanemap from image sensor.