Context
A robot designed to support people with visual impairments needed to identify relevant people and objects using Raspberry Pi-based embedded hardware. The resulting functional embedded solution was taken to production.
All case studies
CASE / 06
Computer Vision · Edge AI · Robotics
Edge AI & Computer Vision
A production embedded-vision pipeline connecting object detection, tracking and prioritization to audio alerts for an assistive robot.
AT A GLANCE
- Role
- Computer Vision Specialist
- Company / context
- Vixsystem · Production assistive robotics system
- Domain
- Computer Vision · Edge AI · Robotics
- Core technology
- Raspberry Pi · ROS · TensorFlow Lite · MobileNet
- Focus
- Detection · Tracking · Audio alerts
01
CONTEXT & PROBLEM
The challenge
The perception flow had to combine detection, tracking and prioritization under embedded-compute constraints, then translate the result into useful audio feedback.
02
MY ROLE
My responsibilities and contribution
I developed the embedded computer-vision flow, evaluated convolutional networks for the hardware constraints and implemented object prioritization for the alert path.
- 01
Developed the computer-vision flow for the assistive robotic system.
- 02
Evaluated convolutional neural networks for execution on embedded hardware.
- 03
Used TensorFlow Lite and MobileNet in the Raspberry Pi-based implementation.
- 04
Implemented object tracking, priority and position logic for audio-alert generation.
03
CONCEPTUAL PERCEPTION-TO-ALERT FLOW
- 01SourceCamera
- 02ConnectRaspberry Pi · ROS
- 03ProcessTFLite · MobileNet
- 04DataDetect · track
- 05ProcessPriority · position
- 06OutcomeAudio alert
Conceptual and anonymized view; implementation details are intentionally omitted.
04
ENGINEERING APPROACH
How the problem was approached
The engineering trade-off was shaped by the physical system: model choice and processing steps had to fit the embedded target while still producing information the alert logic could use.
- 01
Evaluated convolutional-network options against the embedded target.
- 02
Used MobileNet with TensorFlow Lite for the final Raspberry Pi-based flow.
- 03
Connected perception stages through ROS.
- 04
Applied tracking, object priority and position before generating audio alerts.
ENGINEERING CHALLENGES
Challenge → why it mattered → responseEmbedded compute constraints
- Why it mattered
- The vision model needed to execute on a Raspberry Pi-based system.
- Response
- Convolutional networks were evaluated and MobileNet with TensorFlow Lite was used in the implementation.
From detection to useful feedback
- Why it mattered
- Detecting an object alone did not determine which information should become an alert.
- Response
- Tracking, object priority and position were used before the audio-alert stage.
05
SOLUTION
A Raspberry Pi and ROS vision flow used TensorFlow Lite and MobileNet for detection, then applied tracking and prioritization before producing audio alerts.
CORE TECHNOLOGY
Raspberry PiROSTensorFlow LiteMobileNetComputer Vision
06
EVIDENCE-BASED IMPACT
What changed as a result
Recorded outcomes from the project scope.
01
Delivered a functional embedded computer-vision solution that was taken to production.
02
Connected embedded visual perception to prioritized audio feedback in an assistive robotic system.
03
Combined detection, tracking, prioritization and alert generation in one edge-computing flow.