How are you synchronizing multiple cameras and sensors on NVIDIA Jetson Orin?

Hi everyone,

I’m interested in how developers are handling multi-sensor synchronization on NVIDIA Jetson Orin, particularly in robotics and embedded vision applications.

When a system uses multiple cameras alongside sensors such as IMUs or LiDAR, aligning the data can become challenging. Even if each sensor is working correctly, timing differences between streams can affect sensor fusion, object tracking and 3D perception.

Some aspects worth considering include:

  • Hardware timestamps versus software-generated timestamps

  • Synchronizing camera frames with IMU or LiDAR measurements

  • Trigger-based synchronization versus timestamp-based alignment

  • Handling timestamp consistency across multiple sensor streams

  • Managing synchronization in ROS 2 and other perception pipelines

Hardware timestamping is one approach to improving temporal alignment, but the implementation depends on the sensors, drivers and synchronization capabilities of the complete system.

We recently explored an approach to multi-sensor synchronization on NVIDIA Jetson Orin using hardware timestamping, including how it can help align data from different sensors.

Technical article:

For those working with multi-camera or multi-sensor systems on Jetson, I’m curious:

Are you using hardware triggering, hardware timestamps, software timestamps, or a combination of these approaches?

What has been the biggest challenge in getting reliable synchronization across your sensor streams?

On Jetson AGX Orin, reliable multi-sensor sync is a combination of hardware triggering and hardware timestamps. Software timestamps (clock_gettime, V4L2 dqbuf time, ROS header.stamp = now()) are not enough for fusion.

1. Camera-to-camera (exposure alignment)
Use a shared HW FSYNC from the Orin TSC generator (tsc_sig_gen / TSC_EDGE_OUT) into each sensor’s XVS/FSYNC pin, or into the GMSL deserializer FSYNC GPIO and through the serializer. Put every camera in external-trigger / slave mode so they start exposure on the same edge. Free-running sensors with later timestamp matching will still have exposure skew.

FSYNC DT details are in this thread. fsync-groups must be enabled, and start_time must be a TSC tick in the future.

2. Camera timestamps (TSC)
VI latches SOF/EOF in the SoC TSC domain. In Argus use ISensorTimestampTsc:

  • getSensorSofTimestampTsc() — start of frame
  • getSensorEofTimestampTsc() — end of frame

Compare those values across cameras. They are not CLOCK_MONOTONIC / CLOCK_REALTIME.

3. Camera + IMU / LiDAR (one time domain)
TSC and PTP/PHC are different clocks unless they are locked.

  • Ethernet LiDAR / PTP IMU: use NvPPS so TSC locks to the PTP-client MAC (1 Hz PPS). That keeps a fixed offset for fusion.
  • On-board IMU on a kernel/system clock: convert through the NvPPS TSC↔PTP/system mapping. Do not subtract a TSC SOF from a CLOCK_REALTIME IMU sample.

4. ROS 2
Keep the original TSC (or PTP) timestamp in the message. Convert to ROS time only with a known offset from NvPPS. Filling header.stamp at callback time breaks the HW timeline.

Typical failures we see: FSYNC generator left disabled in DT, GMSL GPIO not carrying FSYNC, sensors still rolling-shutter free-run, and mixing Argus TSC with ROS/system time.

We have been using Jetson boards together with PCIe cameras for 9 years. The cameras write their image data into system ram via busmaster DMA, and there is no interface that has more bandwith and less latency then PCIe. Each camera (daylight color, night vision, ir,…) has got its own Jetson board, and all are synced by camera hardware trigger signals.