Bayesian Methods for Parameter Estimation with Event Cameras
This project was done as the final project for the course AE567: Statistical Inference, Estimation and Learning at the University of Michigan. I chose this project as my proposal as it allowed me to use Bayesian statistics to interpret data from a novel type of sensor that I could see myself using in further research and academic applications.
I particularly was interested in using a probabilistic representation of the cameras motion parameters in order to provide an explainable interpretation of the cameras ego-motion.
Event Cameras
Event cameras are sensors that register pixel intensity changes in the image frame. So rather than traditional cameras, which capture frames at a fixed rate, event cameras will asynchronously capture any sort of pixel intensity change. There are a multitude of benefits for such a camera:
Low latency (order of microseconds)
Prevents the camera to suffer from motion blur
High dynamic range
Asynchronous
This provides an extremely dense representation of the motion of the objects in front of the camera.
Why motion parameter estimation?
One of the challenges of event cameras are that because only pixel intensity changes are being tracked, it can be difficult to understand the perspective of the camera, or even how the output data can be interpreted. This is where the task of contrast maximization comes in. By aiming to find the highest contrast output of the event camera, an interpretable representation of the scene in front of the camera can be reconstructed.
Example of contrast maximization on event camera output, provided by Gallergo et al.
It is thankfully intuitive to know how to get the highest contrast image: follow the motion of the camera and adjust the perspective accordingly online!
Example of a high contrast image when following a predicted trajectory from the project.
Report for final project.