Related Experiment Videos
Enhancement of ELDA Tracker Based on CNN Features and Adaptive Model Update
Changxin Gao1, Huizhang Shi2, Jin-Gang Yu3
1National Key Laboratory of Science and Technology on Multispectral Information Processing, School of Automation, Huazhong University of Science and Technology, Wuhan 430074, China. cgao@hust.edu.cn.
Sensors (Basel, Switzerland)
|April 20, 2016
Summary
This study enhances object tracking using deep convolutional neural network (CNN) features and adaptive models. The improved algorithm offers robust and effective visual tracking for complex video scenes.
Area of Science:
- Computer Vision
- Machine Learning
Background:
- Robust visual tracking relies on appearance representation and observation models.
- Exemplar-based linear discriminant analysis (ELDA) shows promise in object tracking.
Purpose of the Study:
- To improve the ELDA tracking algorithm using deep convolutional neural network (CNN) features and adaptive model updates.
- To address the computational cost of CNN feature extraction for real-time tracking.
Main Methods:
- A two-step CNN feature extraction method is proposed, separating convolutional and fully-connected layer computations.
- Both object and background models are updated to enhance adaptivity, balancing discriminative ability and robustness.
- A Gaussian mixture model (GMM) is used for the background model, initialized offline and updated online.
Main Results:
- The proposed tracker was evaluated on a benchmark dataset of 50 challenging video sequences.
- It achieved superior overall performance compared to existing state-of-the-art trackers.
- The results demonstrate the effectiveness and robustness of the enhanced tracking algorithm.
Conclusions:
- The integration of deep CNN features and adaptive model updates significantly improves object tracking performance.
- The proposed two-step feature extraction method efficiently utilizes CNN features.
- The adaptive object and background model updates contribute to robust tracking in complex scenarios.