Why hybrid inference
Small on-device models are dependable without a network. Remote inference makes larger segmentation models practical when the connection and latency budget allow it.
Project 06 · Master's project
Autonomous Driving with Object Segmentation
A small autonomous vehicle used to test which perception tasks can run on a Raspberry Pi and which are better served by a remote GPU.
- Completed
- 2024
- Edge device
- Raspberry Pi 4B
- Models
- FastSAM · YOLOv8-seg
- Measured output
- 20–30 FPS segmentation
Project poster
The system combines local control with networked perception.
Road boundaries and vehicle control stay close to the hardware. Larger segmentation models run externally and return scene information.
System split
Latency, model size, and reliability determine where inference belongs.
- Sense locallyCamera and ultrasonic sensors feed the Raspberry Pi.
- Track the roadCanny edges and Hough lines estimate the vehicle's path.
- Segment remotelyFrames stream to a larger FastSAM or YOLO segmentation model.
- Return decisionsRelevant scene information comes back to the embedded controller.
Engineering scope
The work covered perception, motor control, embedded hardware, networking, and the integration problems between them.
What I learned
Model accuracy is only one system variable. Frame rate, round-trip latency, failure behavior, and deployment complexity decide whether the full product works.
View the implementation ↗