Skip to content

FPV camera and video-system design

An FPV camera is the first part of a complete video system. A good design must consider: - motion distortion - latency - field of view - low-light performance - encoding - radio bandwidth - power noise - mechanical mounting - sensor resolution.

Camera shutter types

The shutter describes how the sensor exposes its pixels.

Rolling shutter

A rolling-shutter sensor exposes and reads the image one row after another. Different rows therefore represent slightly different moments in time.

Advantages:

  • Common, inexpensive sensors
  • High resolution
  • Good image quality and dynamic range
  • Many compact camera modules available

Disadvantages for FPV:

  • Fast rotation can bend vertical objects.
  • Propellers can appear curved or broken.
  • Motor vibration can produce a wobbling jello image.
  • Longer exposure in low light increases motion blur.

Rolling shutter is normally acceptable for general FPV when the camera is mounted rigidly, vibration is controlled, and exposure time remains short.

Global shutter

A global-shutter sensor exposes all pixels during the same time interval. It greatly reduces the geometric distortion created by rapid motion and vibration.

Advantages:

  • Better representation of fast motion
  • Less rolling-shutter wobble and skew
  • Useful for visual odometry, machine vision, and synchronized cameras

Disadvantages:

  • Often lower resolution or more expensive
  • May require a larger module and external lens
  • Does not eliminate ordinary motion blur; exposure time still matters

Raspberry Pi explains the difference using rotating propellers: a rolling shutter scans them line by line, while its Global Shutter Camera captures all pixels together. Raspberry Pi camera documentation

Mechanical shutter

A mechanical shutter physically blocks and exposes the sensor. It is common in still cameras but uncommon in FPV because it adds weight, moving parts, delay, and wear. FPV video systems normally use electronic rolling or global shutters.

Video: global shutter versus rolling shutter

Raspberry Pi camera comparison

The frame rates below are maximum documented sensor/video modes, not guaranteed end-to-end FPV display rates. Encoding, the Raspberry Pi model, radio link, and receiver can reduce the delivered frame rate.

Camera Sensor Shutter Full resolution Example maximum video modes
Camera Module 2 Sony IMX219 Rolling 3280 × 2464, 8 MP 1080p47; 1640 × 1232p41; 640 × 480p206
Camera Module 3 Sony IMX708 Rolling 4608 × 2592, 11.9 MP 2304 × 1296p56; HDR p30; 1536 × 864p120
Camera Module 3 Wide Sony IMX708 Rolling 4608 × 2592, 11.9 MP 2304 × 1296p56; HDR p30; 1536 × 864p120
High Quality Camera Sony IMX477 Rolling 4056 × 3040, 12.3 MP 2028 × 1080p50; 2028 × 1520p40; 1332 × 990p120
AI Camera Sony IMX500 Rolling 4056 × 3040, 12.3 MP 2028 × 1520p30; full resolution p10
Global Shutter Camera Sony IMX296 Global 1456 × 1088, 1.58 MP 1456 × 1088p60

These specifications come from Raspberry Pi's camera hardware comparison. The AI Camera product brief also specifies 30 fps in its 2×2-binned mode and 10 fps at full resolution.

Which camera fits which job?

Goal Good starting option Reason
General digital FPV Camera Module 3 Wide Compact, autofocus, wide view, and useful high-frame-rate modes
Lowest geometric motion distortion Global Shutter Camera Exposes all pixels together
Interchangeable lenses HQ or Global Shutter Camera C/CS-mount lens support
Onboard object detection AI Camera IMX500 includes neural-network acceleration
Low-cost experiment Camera Module 2 Small and widely supported

The AI Camera can produce image and inference streams, but onboard AI does not automatically reduce pilot-view latency. Choose it when inference is part of the mission, not merely because it is newer. Raspberry Pi AI Camera documentation

End-to-end latency

FPV latency is the sum of every pipeline stage:

flowchart LR
    A[Exposure and sensor readout] --> B[ISP processing]
    B --> C[Video encoder]
    C --> D[Transmit queue]
    D --> E[Radio link]
    E --> F[Receiver buffer]
    F --> G[Video decoder]
    G --> H[Display refresh]

Measure glass-to-glass latency—from a real-world event in front of the camera until that event appears on the display. Camera FPS alone does not describe this delay.

At 60 fps, one frame lasts approximately:

\[ \frac{1}{60} = 16.7\text{ ms} \]

Every extra queued frame can therefore add about 16.7 ms before considering encoding, radio, decoding, and display delay.

For low latency:

  • Avoid unnecessary frame queues.
  • Use a hardware encoder when available.
  • Disable B-frames when the encoder supports a low-latency mode.
  • Use a short GOP when quick recovery matters, accepting the bitrate cost.
  • Configure receiver buffers deliberately instead of accepting large defaults.
  • Measure under packet loss, not only on a workbench.

Resolution, frame rate, and bandwidth

Resolution and FPS increase both image detail and required processing and radio bandwidth. Uncompressed BGR video requires approximately:

\[ bitrate = width \times height \times 3 \times FPS \times 8 \]

For 1920 × 1080 at 60 fps:

\[ 1920 \times 1080 \times 3 \times 60 \times 8 \approx 2.99\text{ Gbit/s} \]

An FPV radio link therefore normally sends compressed video such as H.264 or H.265. Actual compressed bitrate depends strongly on motion, scene detail, encoder settings, and acceptable quality.

Choose the operating mode for the complete link. A stable 720p60 stream can be more useful for piloting than a delayed or frequently corrupted 4K stream.

Field of view and lens

A wider lens shows more of the environment and makes rapid rotation feel less disorienting, but distant objects become smaller and edge distortion increases. A narrow lens shows more detail ahead but provides less peripheral awareness.

Check:

  • Horizontal, vertical, and diagonal field of view
  • Lens focal length
  • Fixed focus versus autofocus
  • Minimum focus distance
  • Aperture or f-number
  • Replaceable-lens support
  • Whether the frame or propellers appear in the image

Autofocus can hunt during rapid scene changes. A fixed focus, or locked autofocus position, may provide more predictable FPV behavior.

Exposure, low light, and dynamic range

Short exposure reduces motion blur but collects less light. Increasing analog gain brightens the image but adds noise. For outdoor flight, the camera must also handle bright sky and dark ground in the same frame.

Evaluate:

  • Sensor and pixel size
  • Lens aperture
  • Maximum useful gain
  • Minimum practical exposure time
  • HDR or wide-dynamic-range mode
  • Highlight recovery and shadow noise
  • Daylight, sunset, and indoor behavior

HDR may improve dynamic range but can reduce maximum frame rate or add processing latency. Test the exact mode used by the video link.

Camera interface and processing hardware

Raspberry Pi cameras normally use CSI-2. Confirm:

  • Camera connector type and ribbon-cable compatibility
  • Number of CSI lanes and supported sensor mode
  • Raspberry Pi model and hardware-encoder capabilities
  • ISP capacity when capturing multiple streams
  • Memory bandwidth and CPU/GPU load
  • Software support in libcamera, rpicam-apps, or Picamera2
  • Whether recording, AI inference, and streaming can run together at the target FPS

Do not assume that a sensor's maximum mode can also be encoded, transmitted, decoded, and displayed at that rate.

Power and electrical noise

Motors and ESCs create substantial switching noise. A camera or transmitter powered from a noisy rail can show horizontal lines, resets, or corrupted data.

Plan for:

  • Correct input-voltage range
  • Peak and average power consumption
  • A regulator with enough current and thermal margin
  • Input and local decoupling capacitors
  • Short power and ground paths
  • Separation from ESC phase wires and high-current battery wiring
  • Ground-loop avoidance
  • Filtering validated under full motor load