Congestion Analysis of Convolutional Neural Network-Based Pedestrian Counting Methods on Helicopter Footage

TL;DR

This paper evaluates how well current CNN-based pedestrian counting methods perform on aerial helicopter footage, especially under different congestion levels, highlighting challenges and differences from street-level datasets.

Contribution

It provides a comprehensive analysis of pedestrian counting performance on aerial imagery and examines the impact of congestion levels, addressing a gap in existing research.

Findings

Performance decreases with higher congestion levels

Aerial images pose unique challenges compared to street-level datasets

State-of-the-art methods vary significantly in accuracy on helicopter footage

Abstract

Over the past few years, researchers have presented many different applications for convolutional neural networks, including those for the detection and recognition of objects from images. The desire to understand our own nature has always been an important motivation for research. Thus, the visual recognition of humans is among the most important issues facing machine learning today. Most solutions for this task have been developed and tested by using several publicly available datasets. These datasets typically contain images taken from street-level closed-circuit television cameras offering a low-angle view. There are major differences between such images and those taken from the sky. In addition, aerial images are often very congested, containing hundreds of targets. These factors may have significant impact on the quality of the results. In this paper, we investigate state-of-the-art methods for counting pedestrians and the related performance of aerial footage. Furthermore, we analyze this performance with respect to the congestion levels of the images.