Algorithm Description

This code includes the C++ implementation of the Ensemble of Kernelized Correlation Filter Tracker (EnKCF). The EnKCF runs multiple KCFs [1] to tackle different aspects of tracking such as : scale, and fast motion. It also employs a Particle Filter to smoothen the interaction among different KCFs. Our tracker achieves higher success and precision rates than the baseline tracker at 416hz on UAV123 dataset. Below, you can find the hyper-parameters and their optimal values. The EnKCF is inspired by the long-term correlation (LCT) tracker [2], however, our goal is to utilize multiple KCFs in an efficient way to keep the complexity at each frame similar to the baseline KCF (O(nlogn)) [1]. This way, it maintains (30fps) operation rate on computationally limited embedded systems.

You can find more details on the EnKCF tracker in our arxiv version of the paper.

EnKCF Hyperparameters

• sigma_scale = 0.9 // Gaussian Kernel Bandwith in Scale KCF
• sigma_large_roi_translation = 0.7 // Gaussian Kernel Bandwith in Large ROI Trans. KCF
• sigma_small_roi_translation = 0.6 // Gaussian Kernel Bandwith in Small ROI Trans. KCF
• lambda = 0.0001 // Regularization Weight - Same for all the KCFs
• scale_filter_frequency = 5 // Scale Filter Applied every 5 frames
• learning_rate_scale = 0.10 // Make it 0.25 for the UAV123_10fps dataset
• learning_rate_large_roi_translation = 0.20
• learning_rate_small_roi_translation = 0.20
• scale_filter_training_psr_threshold = 4.0 // Threshold to Train Scale Filter
• padding_scale_filter = 0.0 // Area to Consider for Scale Filter
• padding_large_roi_translation = 2.0 // Area to Consider for Large Area Trans. Filter
• padding_small_roi_translation = 1.5 // Area to COnsider for Small Area Trans. Filter
• responsevariance_scale = 0.04 // Variance for Desired Gaussian Response for Scale KCF
• responsevariance_large_roi_translation = 0.06 // Variance of the Gaussian Response for Large ROI Translation KCF
• responsevariance_small_roi_translation = 0.125 // Variance of the Gaussian Response for Small ROI Translation KCF

Particle Filter Hyperparameters

It should be highlighted that we recommend that the particle filter to be removed in the existence of global camera motion.

• number_particles = 300 // Number of Particles in the Particle Filter
• number_efficient_particles = 1000/3.0 // Number of Efficient Particles to Enable Resampling
• process_noise_uniform_x = [-10,10] // Transition Noise X Coordinate
• process_noise_uniform_y = [-10,10] // Transition Noise Y Coordinate
• process_noise_uniform_vx = [-2,2] // Transition Noise X Velocity
• process_noise_uniform_vy = [-2,2] // Transition Noise Y Velocity
• transition_noise_uniform_x = [-25,25] // Distribution Interval X Coordinate
• transition_noise_uniform_y = [-25,25] // Y Coordinate
• transition_noise_uniform_vx = [-10,10] // X Velocity
• transition_noise_uniform_vy = [-10,10] // Y Velocity
• beta_weight_function = 0.05 // exp(-dist * beta) - Weight Function Hyperparameter - For Importance Sampling based on spatial Eclidean Distance to Maximum of response map

Run the Tracker

More information on running the tracker can be found here.

For your questions or comments, please contact Burak Uzkent at [email protected].

MATLAB code for the EnKCF tracker can be found here. A README file for the MATLAB code will be added soon.

You can cite the EnKCF tracker as

B. Uzkent and Y. Seo, "EnKCF: Ensemble of Kernelized Correlation Filters for High-Speed Object Tracking," 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, 2018, pp. 1133-1141.
doi: 10.1109/WACV.2018.00129

References

[1] - Henriques, João F., Rui Caseiro, Pedro Martins, and Jorge Batista. "High-speed tracking with kernelized correlation filters." IEEE Transactions on Pattern Analysis and Machine Intelligence 37, no. 3 (2015): 583-596.

[2] - Ma, Chao, Xiaokang Yang, Chongyang Zhang, and Ming-Hsuan Yang. "Long-term correlation tracking." In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 5388-5396. 2015.