This is a research project for ROB 590 Independent Study for Winter 2022 at the University of Michigan. The goal of this project is to perform Multiple Target Tracking for Event Camera Data. The work has been evaluated on 3 indoor datasets collected by the research group at the University of Zurich, for target tracking using event streams.
Clone the repository.
git clone [email protected]:rastri-dey/Event_Camera_MTT.git
mkdir Data
Download the dataset from https://rpg.ifi.uzh.ch/davis_data.html. The dataset are available in zip format of text files. Convert the .txt files in MATLAB table format and save as .mat files indicated below:
Save "shapes_rotation.txt" as "events_shapes_rotation.mat"
Save "shapes_translation.txt" as "events_shapes_translation.mat"
Save "shapes_6dof.txt" as "events_shapes_6dof.mat"
Please note, the mat files should be in table format to read the event data from the main script. Run the main file.
Event_Camera_Multiple_Target_Tracking.m
The sample results are shown here. Detailed results for trajectory with respect to ground truth for all the sequences are present in \images folder.
The target trajectory for the 3 datasets has been evaluated and compared with the existing approach by the research group: https://ieeexplore.ieee.org/document/8593380?msclkid=a5bfd524bff011ec9414f0f574b704cf. Under all the 3 datasets from https://rpg.ifi.uzh.ch/davis_data.html, the event based tracking performs with high accuracy. The table shows the mean error of the target tracking for 4 objects in 3 different sequences executing varying motion dynamics in X and Y pixel coordinates.
This research would not have been possible without the exceptional support of Professor Katie Skinner at the University of Michigan, Ann Arbor.