Learning Complexity-Aware Cascades for Pedestrian Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian detection is crucial for autonomous systems.
• Existing cascade detectors struggle with integrating diverse feature complexities.
• Deep convolutional neural networks (CNNs) offer powerful features but are computationally intensive.

Purpose of the Study:

• To investigate complexity-aware cascaded pedestrian detectors.
• To introduce a novel cascade design procedure balancing accuracy and complexity.
• To enable the integration of complex features, like CNN responses, into cascade detectors.

Main Methods:

• Formulating cascade learning as Lagrangian optimization of a risk function.
• Developing a boosting algorithm named complexity-aware cascade training (CompACT).
• Designing pedestrian detectors using features with vastly different complexities, including CNNs.

Main Results:

• CompACT cascades achieve an optimal trade-off between accuracy and complexity.
• Higher complexity features are effectively utilized in later cascade stages.
• Pedestrian detectors integrating CNNs seamlessly achieved accurate and fast detection.

Conclusions:

• CompACT provides an effective method for designing efficient and accurate pedestrian detectors.
• The approach allows for the seamless integration of deep learning features within cascade architectures.
• This work advances pedestrian detection by enabling the use of a wider range of features.