Real-Time Human Detection for Aerial Captured Video Sequences via Deep Models.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional human detection methods rely on handcrafted features, limiting their adaptability to dynamic conditions like illumination changes and camera movement.
• Feature learning approaches offer automated, abstract, and discriminative feature extraction, reducing reliance on expert knowledge and improving efficiency.

Purpose of the Study:

• To evaluate automatic feature learning methods for human detection in aerial videos using non-static cameras.
• To compare the performance, accuracy, and learning speed of three deep learning models: supervised Convolutional Neural Network (S-CNN), pretrained CNN feature extractor, and Hierarchical Extreme Learning Machine (H-ELM).

Main Methods:

• Combined optical flow with three deep learning models: S-CNN, pretrained CNN, and H-ELM.
• Trained and tested models on the challenging UCF-ARG aerial dataset, considering five human actions: digging, waving, throwing, walking, and running.
• Analyzed training/testing accuracy and learning speed on both Central Processing Unit (CPU) and Graphical Processing Unit (GPU).

Main Results:

• Pretrained CNN achieved the highest average accuracy (98.09%).
• S-CNN achieved 95.6% (with softmax) and 91.7% (with SVM) accuracy.
• H-ELM achieved 95.9% accuracy with a training time of 445 seconds on a CPU, while S-CNN training took 770 seconds on a GPU.

Conclusions:

• Automatic feature learning methods, particularly pretrained CNN, are highly effective for human detection in challenging aerial video datasets.
• The study validates the success of deep learning models in recognizing human actions from aerial platforms with dynamic camera conditions.
• Comparative analysis provides insights into the trade-offs between accuracy and learning speed for different deep learning architectures in human detection tasks.