Estimate the vehicle motion based on vision cue: optic flow and pose estimation
1.Self-driving as a promising new technology 2.Developments made over the years 3.Existing various optical flow estimation algorithm
Currently there are many algorithms regarding optical flow in self driving. We want to improve classical methods so that it is more resistant to noise caused by weather and other environmental issues.
1.The total dataset is available by group 20. The demo datasets is available here
There is an example data segment in this repo for experimentation. There are also some notebooks with some example code. Including a position benchmark. Make sure to pip install -r requirements.txt
if you do not have the relevant packages installed already.
The examples contain a 1 minute sample segment and some sample notebooks.
- Lucas-Kanade naive method implementation
- Pyramidal Lucas-Kanade Optical Flow overlay on CNN architecture -- 3 Dimensional parameter into CNN
- Sublinear Optical Flow Algorithm on CNN architecture -- 3 Dimensional parameter into CNN
- Five image cue differences on CNN architecture -- 5 Dimensional parameter into CNN
Structure The ideas is to split the video sources into individual picture frame and use csv file to store its speed at a particular speed. csv-files | | +-- images-paths | speed@frame
1.Forward wrapping 2.Lucas-Kanade Optical Flow
-
Pyrimid Lucas-Kanade Optical Flow overlay -- 3 Dimensional parameter
-
Sublinear Optical Flow Algorithm -- 3 Dimensional parameter
-
Five image cue differences -- 5 Dimensional parameter
The network has about 27 million connections and 250 thousand parameters.
Thanks to https://www.digitaltrends.com/cars/commaai-self-driving-retrofit-software-free/