Mask-Guided Attention Network and Occlusion-Sensitive Hard Example Mining for Occluded Pedestrian Detection

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep convolution neural networks (CNNs) have advanced pedestrian detection.
• Performance on heavily occluded pedestrians remains a significant challenge.
• Occlusions arise from intra-class (other pedestrians) and inter-class (cars, bicycles) interactions.

Purpose of the Study:

• To develop an effective approach for detecting heavily occluded pedestrians.
• To improve the accuracy and robustness of pedestrian detection systems in complex urban environments.
• To establish a new state-of-the-art in occluded pedestrian detection.

Main Methods:

• Introduction of a novel mask-guided attention network to emphasize visible pedestrian regions and suppress occluded ones.
• Development of an occlusion-sensitive hard example mining method and loss function to prioritize difficult, occluded cases.
• Empirical validation using weak box-based segmentation annotations as an approximation for dense pixel-wise annotations.

Main Results:

• The proposed approach achieves state-of-the-art performance on the CityPersons, Caltech, and ETH datasets.
• A significant absolute gain of 10.3% in log-average miss rate was achieved on the heavily occluded HO pedestrian set of the CityPersons test set.
• The method demonstrates effectiveness in handling diverse occlusion patterns.

Conclusions:

• The mask-guided attention network effectively addresses the challenge of occluded pedestrian detection.
• The occlusion-sensitive mining and loss strategies enhance model performance on difficult samples.
• The approach offers a robust and accurate solution for real-world pedestrian detection systems.