Flow on the Go

Ashwin Sekar (asekar) and Richard Zhao (richardz)

Summary

We implement real time optical flows on a mobile GPU platform using the dense inverse search method.

Background

A common problem in computer vision is detecting moving objects on a background. With an increasing amount of cameras mounted on moving vehicles, stabilization of the video feed is a crucial preprocessing task.

Optical flows present an elegant solution to a wide class of problems such as the above. An optical flow is a vector field that describes per-pixel displacements between two consecutive video frames in a video feed.

In recent years, there has been increased interest in algorithms for computing optical flows, especially ones that achieve a mix of efficiency and accuracy. Kroeger et. al. propose a method with very low time complexity and competitive accuracy for computing dense optical flow^[1].

The algorithm is highly parallelizable, which gives it the potential to achieve super-real-time (faster than 30 Hz) performance on GPUs.

Build

Reference

make flow_ref

Resources

[1] Tim Kroeger, et. al Fast Optical Flow using Dense Inverse Search (2016)

Schedule