Lead Embedded AI Engineer – Sports Motion Capture & Analysis

Location: [Remote / On-site] - China
Type: Full-time / Contract 

About the Role

We are building a next-generation on-device AI coaching system that captures human movement (full body, legs, arms, and joint angles) using visual and positional sensors. Unlike cloud-reliant solutions, our device processes data locally on lightweight hardware (e.g., Raspberry Pi / edge SoCs) to deliver real-time, low-latency feedback. We are seeking a hybrid hardware-software engineer who can own the entire pipeline—from sensor selection and circuit integration to embedded Linux optimization and AI model deployment.

Key Responsibilities

• Hardware Architecture & Selection

  • Specify and integrate imaging components: high-speed cameras (e.g., CSI MIPI modules, global-shutter sensors), GPS/GNSS receivers, and optional IMUs for sensor fusion.

  • Evaluate and select single-board computers or SoMs (e.g., Raspberry Pi CM5, NVIDIA Jetson Nano, or Rockchip RK3588) based on compute-per-watt and thermal constraints.

  • Design or adapt carrier boards / power regulation for field-portable operation (battery-powered, ruggedized).

• Software & Embedded Systems

  • Build an embedded Linux pipeline (Yocto/Buildroot or Raspberry Pi OS) that handles camera streaming, GPS parsing, and sensor synchronization.

  • Implement low-latency image acquisition using V4L2, GStreamer, or libcamera, with frame capture rates ≥ 30 fps.

  • Write C++/Python middleware to time-stamp and fuse video frames with GPS/IMU data for spatial-temporal alignment.

• Lightweight AI & Computer Vision

  • Deploy on-device pose estimation using open-source frameworks: MediaPipe, OpenPose, YOLOv8n-pose, or ViTPose (quantized/TFLite versions).

  • Optimize models for edge hardware using TensorFlow Lite, ONNX Runtime, or NPU/GPU acceleration (e.g., Vulkan, Coral Edge TPU).

  • Implement skeleton extraction (33+ keypoints) and joint-angle calculations (e.g., elbow/knee flexion) against ideal reference postures.

• Data Pipeline & Feedback

  • Design a real-time streaming architecture to compare captured motion against pre-loaded "ideal" templates (e.g., golf swing, running gait, tennis serve).

  • Build a visual/audio feedback loop (e.g., LED indicators, buzzer, or simple UI overlay) for live coaching corrections—all processed on-device, with < 100 ms end-to-end latency.

• Testing & Calibration

  • Develop calibration routines for camera intrinsics/extrinsics and GPS-to-world coordinate mapping.

  • Validate system accuracy against Vicon/Qualisys-grade motion capture (or synthetic ground truth).

Required Qualifications

• Education: B.S./M.S. in Electrical Engineering, Computer Engineering, Robotics, or related field.

• Hardware Experience:

  • Hands-on with camera modules (IMX219/IMX477, OV9281, etc.) and MIPI CSI-2 interfaces.

  • Familiarity with GNSS receivers (u-blox, NEO-series) and NMEA parsing.

  • Experience selecting power supplies, batteries, and thermal management for portable devices.

• Software & AI Experience:

  • Proficient in Python and C++ (embedded environment).

  • Deployed at least one pose estimation model on an edge device (Raspberry Pi / Jetson).

  • Strong understanding of model quantization (INT8/FP16), pruning, and frame-skipping strategies to meet real-time constraints.

• Open-Source Proficiency:

  • Deep familiarity with OpenCV, MediaPipe, TensorFlow Lite, and GStreamer.

  • Experience with ROS/ROS2 or custom sensor-fusion frameworks is a plus.

• Systems Thinking: Ability to make build-vs-buy decisions on computing resources, balance cost vs. performance, and document trade-offs.

Nice-to-Have

• Experience with FPGA or Coral Edge TPU for ultra-low-power acceleration.

• Knowledge of sports biomechanics or kinesiology (joint angle conventions, sagittal/frontal planes).

• Background in wireless streaming (Wi-Fi/BLE) for optional cloud backup.

• Familiarity with Docker and CI/CD for edge device OTA updates.