Skip to content
This repo contains MATLAB implementations of the event-based feature tracking methods described in "Event-based Feature Tracking with Probabilistic Data Association" and "Event-based Visual Inertial Odometry".
Branch: master
Clone or download
alexzzhu Merge pull request #4 from ialzugaray/master
Fix IntrinsicMatrix format in cameraParameters
Latest commit 126ed5e Nov 23, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
EventFeatureTracking Fixed input arguments to em2_affine May 3, 2018
data Fix IntrinsicMatrix format Nov 19, 2018
.gitignore Initial commit Mar 25, 2018 Initial commit Mar 25, 2018 Update Mar 27, 2018

Event Feature Tracking

This repo contains MATLAB implementations of the event-based feature tracking methods described in Event-based Feature Tracking with Probabilistic Data Association and Event-based Visual Inertial Odometry. The code has been tested to run on MATLAB 2017b.


We provide functions in the data folder to convert event data from ROS bag to MATLAB mat files. Currently, data from the Event Camera Dataset and the Multi Vehicle Stereo Event Camera dataset is tested. The extract_ros_data.m and either the generate_undistort_map_radtan.m or functions are necessary to generate the data and undistort map mat files. You must select the appropriate undistort map function depending on whether the camera calibration uses the radtan distortion model (e.g. the Event Camera Dataset) or the equidistant distortion model (e.g. MVSEC). Note that you will need the matlab_rosbag package, which has pre-compiled releases here. Assuming you have the boxes_6dof.bag file in the data folder, the extraction code can be run as follows:

load_folder = 'data';
save_folder = 'data';
extract_ros_data('boxes_6dof', load_folder, save_folder)

For the radtan distortion model, the undistort maps are generated in MATLAB as follows:

load data/boxes_6dof.mat
generate_undistort_map_radtan(cinfo, 'boxes_6dof')

For the equidistant distortion model, the undistort maps are generated from the terminal as follows:

python --camchain data/camchain-imucam-indoor_flying1.yaml

Note that the undistort_map can be shared across sequences with the same camera intrinsics.

Running the code

The code can be run from the script main.m. Before you do so, you must set a few parameters, which are stored in the function get_params.m. At a minimum, you will need to set the path to the generated data mat (params.data_path), as well as the generated undistort_map mat (params.undistort_data_path). This function also provides many other parameters, such as start and end times in the sequence, number of features and many more. For optimal performance, the main tracking loop can be run in parallel by setting params.do_parallel to true.

If you are using the IMU related features (for the EM2 affine warp or two point RANSAC), you will need the Robotics System Toolbox, or to define your own functions for quat2rotm and rotm2quat, which convert between a quaternion (w,x,y,z) and a 3x3 rotation matrix.

The feature points and their corresponding IDs are stored as the variables valid_feature_positions and valid_ids, and are updated at the end of each iteration in main.m. If params.headless is set to false, the features will be plotted on top of the integrated event image in Figure 1.

Note that the label "EM1 flow" refers to the optical flow estimation step (Section IV.B in Event-based Feature Tracking with Probabilistic Data Association), and the label "EM2 affine" refers to the feature alignment step (Section IV.C in Event-based Feature Tracking with Probabilistic Data Association).


To debug the underlying EM algorithms, set params.debug to true. This will plot the optimization process at each step, and may help give you an intuition behind what is being minimized. Note that setting this to true will automatically set params.do_parallel to false, as plotting doesn't work properly inside a parfor loop.


For comparisons and benchmarks, the method proposed in "Event-based Feature Tracking with Probabilistic Data Association" can be run by setting params.use_em1_template, params.use_imu and params.do_ransac to false. Setting these parameters to true will run the full method proposed in "Event-based Visual Inertial Odometry".


If you use this code in an academic publication, please cite the following works:

Alex Zihao Zhu, Nikolay Atanasov and Kostas Daniilidis. "Event-based Feature Tracking with Probabilistic Data Association", IEEE International Conference on Robotics and Automation (ICRA), 2017.

Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. "Event-Based Visual Inertial Odometry." IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 2017.

This technology is the subject of a pending patent application: PCT/US2018/018196

You can’t perform that action at this time.