Vision for Crowd Management
Self-supervised and Unsupervised Learning
January 25, 2024Deep Fake Detection
January 25, 2024
Leveraging advanced vision technologies to optimize crowd analysis and facilitate effective management in diverse scenarios.
a. Abnormal Behavior Recognition in Crowds
This project investigates end-to-end abnormal behavior detection frameworks,
leveraging a combination of GCN and 3D CNN. These approaches employ a one-
class classifier, trained on abnormal scores with GCN to enhance detection by
modeling similarities between video clips.
Related Publications
b Croud Counting
This project addresses the challenges in crowd counting and density map
estimation by proposing multi-task attention-based crowd counting networks.
These networks incorporate density-level classification, density map estimation,
and segmentation-guided attention to enhance accuracy in both sparse and
crowded scenarios.
Related Publications
c. Automatic Augmentation for Crowd Counting
This study introduces an automatic augmentation framework for crowd counting
using deep reinforcement learning to address the challenge of insufficient labeled
training samples. This project aims to employ the deep deterministic policy
gradient for crowd counting.
Related Publications
d. Person Localization in Spectator Crowds
This project introduces multi-person head segmentation algorithm using
convolutional encoder-decoder networks for person head detection in crowded
scenes. For this purpose, probability heatmaps are utilized. The algorithms excel
in assigning probabilities to head pixels and capturing multiresolution
information. It aims to simultaneously detect all heads and faces, that
compensating down-sampling information loss.
Related Publications
e. Action Recognition is Spectator Crowds:
This work aims to identify spectator actions in large crowds for security and saftey
perspective. To address real-world surveillance scenarios, low resolution crowd
videos are considered.
Related Publications