研究目的
To improve infrared pedestrian classification performance by suppressing cluttered background using automatic image matting and deep learning, without increasing the depth of neural networks.
研究成果
The proposed automatic matte-based deep learning approach significantly improves infrared pedestrian classification performance by suppressing cluttered background and providing consistent input to deep neural networks. It achieves better results than state-of-the-art methods on multiple datasets with negligible computational cost increase. However, it is limited by its assumptions and may not perform well in all scenarios, indicating a need for further evaluation in other contexts.
研究不足
The approach assumes that pedestrians walk upright in infrared images and that the head appears brighter than the background. It may fail in complex cases where these assumptions are not satisfied, such as when trimaps are incorrectly generated due to background clutter or changes in posture. Additionally, the method is specific to infrared images and may not generalize well to color or gray images without further investigation.
1：Experimental Design and Method Selection:
The study proposes an automatic matte-based deep learning approach for infrared pedestrian classification. It involves automatic infrared pedestrian matting preprocessing (including automatic pedestrian trimap generation and infrared image matting) and automatic matte-based deep learning using a convolutional neural network (AlexNet). The global matting approach is adopted for image matting, and a hill climbing algorithm is used for precise head location.
2：Sample Selection and Data Sources:
Experiments were conducted on three benchmark datasets: LSI far infrared pedestrian dataset, RIFIR far infrared dataset, and KAIST multispectral pedestrian detection benchmark dataset. Only infrared images were used.
3：List of Experimental Equipment and Materials:
A desktop PC with a 3.1 GHz quad-core Intel Core i5 processor, 8 GB memory, and an Nvidia GTX 1080 GPU for acceleration. The Caffe platform was used for deep learning implementation, and the automatic pedestrian matting was implemented in C++.
4：1 GHz quad-core Intel Core i5 processor, 8 GB memory, and an Nvidia GTX 1080 GPU for acceleration. The Caffe platform was used for deep learning implementation, and the automatic pedestrian matting was implemented in C++.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: For each input infrared image, it is standardized to 32x64 size. Then, automatic trimap generation is performed by locating the pedestrian's head and upper body. Image matting is applied using the global sampling strategy to produce alpha mattes. Finally, AlexNet is used for classification with the alpha mattes as input.
5：Data Analysis Methods:
Performance was evaluated using precision, recall, accuracy, F-measure, and ROC curves. Thresholds were selected to maximize AUC. Comparisons were made with other preprocessing approaches and deep learning methods.
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
