Too Far to See? Not Really!-Pedestrian Detection With Scale-Aware Localization Policy

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian detection systems struggle with small, distant pedestrians due to performance degradation.
• Pedestrians at different distances have varying visual characteristics.

Purpose of the Study:

• To develop an active pedestrian detector that overcomes limitations in detecting small and distant pedestrians.
• To leverage multi-layer feature representations for improved detection accuracy.

Main Methods:

• Utilized convolutional neural networks (CNNs) like ResNet and Faster R-CNNs for hierarchical feature extraction.
• Employed a region proposal network and region of interest pooling for initial pedestrian proposals.
• Developed an active detector that uses coordinate transformations on proposals to identify location and optimal feature layer.

Main Results:

• Achieved lower overall detection errors on standard benchmarks compared to state-of-the-art methods.
• Demonstrated significant improvement for far-scale pedestrians (below 80 pixels height).
• Reduced the log-average miss rate for far-scale pedestrians from 60.51% to 41.85% on the Caltech benchmark.

Conclusions:

• The proposed active pedestrian detector effectively handles variations in pedestrian scale and distance.
• Multi-layer feature analysis combined with an active detection mechanism enhances performance for challenging scenarios.
• This method offers a substantial improvement for real-world pedestrian detection applications.