Counting Crowds with Perspective Distortion Correction via Adaptive Learning.

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Crowd counting aims to estimate the number of people in images, with regression-based and Convolutional Neural Network (CNN)-based methods being prominent.
• Locating individuals within crowded scenes presents challenges due to perspective distortion, which alters crowd size perception.

Purpose of the Study:

• To develop a novel framework, the Adaptive Learning Network (CAL), for accurate crowd counting and person localization.
• To address the challenge of perspective distortion in crowd counting by effectively integrating multi-scale features.

Main Methods:

• The proposed Adaptive Learning Network (CAL) utilizes VGG as its backbone architecture.
• It extracts features at different scales (1/2, 1/4, 1/8, 1/16) after each pooling layer.
• These multi-scale features are combined using weights learned by an adaptive learning branch, tailored to each image in the dataset.

Main Results:

• Experiments were conducted on benchmark datasets: ShanghaiTech Part A, ShanghaiTech Part B, UCF_CC_50, and UCF-QNRF.
• The CAL model demonstrated robust performance in crowd counting and person localization across these diverse datasets.
• The adaptive learning approach effectively reduced the impact of perspective distortion on crowd size estimation.

Conclusions:

• The Adaptive Learning Network (CAL) offers an effective solution for crowd counting, particularly in scenes with significant perspective distortion.
• Integrating multi-scale features through adaptive learning enhances the accuracy of both crowd counting and individual person localization.
• This framework represents a significant advancement in addressing the complexities of crowd analysis in computer vision.