I've started working on a project about the trajectory data mining from videos - for example: snowboarding video from GoPro action camera.
This is the continuation of my previous experiment (MotionML), which was based on the accelerometer datasets and KNN-DTW classification: https://github.com/llvll/motionml
Visual odometry and monocular SLAM topics have been researched with the following algorithms:
At this stage I've selected ORB-SLAM algorithm for the initial experiments: https://github.com/raulmur/ORB_SLAM
The implementation will be done in Python using Scikit-Learn and FastDTW for the model training purposes based on the trajectory data. The data mining itself, based on OpenCV and ORB-SLAM, will execute on Apache Spark for video file processing.
The following results are expected:
- Extracted trajectory data.
- Discovered and classified motion patterns - similar to MotionML project above.
- Video georeferencing or spatio-temporal referencing - GPS geotagging might be present for frame #1, trajectory data will extrapolate geotagging to the next frames. Clock synchronization should take place.
Two questions from my side:
DTW for trajectory classification - I'm still doubting about DTW and reviewing HMM (Hidden Markov Models) or LSTM neural nets for the same purpose. Any practical advices are really appreciated.
I'm also looking for any examples or advices about SLAM-based feature extraction for machine learning. Ideally for trajectory data mining, but other examples will work as well.
If you would like to cooperate or to participate in this project, then please write me directly: firstname.lastname@example.org