# Computer Vision (Optical Flow, MoCap, VIO, Avoidance)

Computer vision (opens new window) techniques enable computers to use visual data to make sense of their environment.

PX4 uses computer vision systems (primarily running on Companion Computers) in order to support the following features:

  • Optical Flow provides 2D velocity estimation (using a downward facing camera and a downward facing distance sensor).
  • Motion Capture provides 3D pose estimation using a vision system that is external to the vehicle. It is primarily used for indoor navigation.
  • Visual Inertial Odometry provides 3D pose and velocity estimation using an onboard vision system and IMU. It is used for navigation when GNSS position information is absent or unreliable.
  • Obstacle Avoidance provides full navigation around obstacles when flying a planned path (currently missions are supported). This uses PX4/PX4-Avoidance (opens new window) running on a companion computer.
  • Collision Prevention is used to stop vehicles before they can crash into an obstacle (primarily when flying in manual modes).

TIP

The PX4 Vision Autonomy Development Kit (Holybro) is a robust and inexpensive kit for developers working with computer vision on PX4. It comes with no pre-installed software, but does include an example implementation of obstacle avoidance to demonstrate the capabilities of the platform.

# Motion Capture

Motion Capture (MoCap) is a technique for estimating the 3D pose (position and orientation) of a vehicle using a positioning mechanism that is external to the vehicle. MoCap systems most commonly detect motion using infrared cameras, but other types of cameras, Lidar, or Ultra Wideband (UWB) may also be used.

Note

MoCap is commonly used to navigate a vehicle in situations where GPS is absent (e.g. indoors), and provides position relative to a local coordinate system.

For information about MoCap see:

# Visual Inertial Odometry (VIO)

Visual Inertial Odometry (VIO) is used for estimating the 3D pose (position and orientation) and velocity of a moving vehicle relative to a local starting position. It is commonly used to navigate a vehicle in situations where GPS is absent (e.g. indoors) or unreliable (e.g. when flying under a bridge).

VIO uses Visual Odometry (opens new window) to estimate vehicle pose from visual information, combined with inertial measurements from an IMU (to correct for errors associated with rapid vehicle movement resulting in poor image capture).

Note

One difference between VIO and MoCap is that VIO cameras/IMU are vehicle-based, and additionally provide velocity information.

For information about configuring VIO on PX4 see:

# Optical Flow

Optical Flow provides 2D velocity estimation (using a downward facing camera and a downward facing distance sensor).

For information about optical flow see:

# Comparisons

# Optical Flow vs VIO for Local Position Estimation

Both these techniques use cameras and measure differences between frames. Optical flow uses a downward facing camera, while VIO uses a stereo camera or a 45 degree tracking camera. Assuming both are well calibrated, which is better for local position estimation?

The consensus appears to be (opens new window):

Optical flow:

  • Downward facing optical flow gives you a planar velocity thats corrected for angular velocity with the gyro.
  • Requires an accurate distance to the ground and assumes a planar surface. Given those conditions it can be just as accurate/reliable as VIO (such as indoor flight)
  • Is more robust than VIO as it has fewer states.
  • Is significantly cheaper and easier to set up as it only requires a flow sensor, a rangefinder, and setting up a few parameters (which can be connected to the flight controller).

VIO:

  • Is more expensive to purchase and harder to set up. It requires a separate companion computer, calibration, software, configuration and so on.
  • Will be less effective if there are no point features to track (in practice the real world generally has point features).
  • Is more flexible, allowing additional features such as obstacle avoidance and mapping.

A combination (fusing both) is probably the most reliable, though not necessary in most real-world scenarios. Normally you will select the system that suits your operating environment, required features, and cost constraints:

  • Use VIO if you plan on flying outdoors without GPS (or outdoors and indoors), or if you need to support obstacle avoidance and other computer vision features.
  • Use Optical Flow if you plan on only flying indoors (without GPS) and cost is an important consideration.

# External Resources