研究目的
To investigate the performance of Kernelized Correlation Filter (KCF) trackers using deep features versus HOG features in thermal image tracking, and to demonstrate that pre-trained deep features from visual object classification are not effective for thermal tracking due to lack of invariance to appearance changes from camera type differences.
研究成果
The KCF algorithm performs better with HOG features than with pre-trained deep features in thermal image tracking for most cases (12 out of 14 videos), indicating that deep features from visual classification networks are not robust to appearance changes in thermal imagery. This suggests a need for improved CNN architectures and training methods specifically adapted for thermal data to enhance correlation filter-based trackers.
研究不足
The study is limited to specific thermal video datasets (LTIR v1.0), and the deep features are from CNNs pre-trained on visual object classification tasks, which may not generalize to other thermal datasets or camera types. The performance comparison is based on a small set of 14 videos, and the results may not hold for all thermal tracking scenarios. Optimization of CNN architectures and training methods for thermal data is not explored in depth.
1：Experimental Design and Method Selection:
The study uses the Kernelized Correlation Filter (KCF) algorithm for object tracking, comparing its performance with Histogram of Oriented Gradients (HOG) features and deep features from pre-trained Convolutional Neural Networks (CNNs). The methodology involves mathematical derivations of correlation filters and kernel tricks for efficient computation.
2：Sample Selection and Data Sources:
Thermal video datasets are selected from the Link?ping Thermal InfraRed (LTIR) dataset v1.0, which includes sequences from different thermal sensors and challenging scenarios like occlusions and appearance changes.
3：0, which includes sequences from different thermal sensors and challenging scenarios like occlusions and appearance changes.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: Thermal cameras (cooled and uncooled types, specific models not mentioned), computers with high-performance GPUs for deep learning computations, and software for implementing KCF and feature extraction.
4：Experimental Procedures and Operational Workflow:
Image patches are sampled from thermal video frames; features (HOG or deep) are extracted; KCF is applied for tracking; precision curves are generated to evaluate tracking accuracy based on distance thresholds.
5：Data Analysis Methods:
Precision curves are used to assess tracking performance, comparing the percentage of correctly tracked frames for HOG and deep features across multiple video sequences.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
