TLDR: VelocityNet is a new framework for real-time anomaly detection in densely crowded environments. It uses a dual-pipeline system combining head detection and optical flow to calculate individual person velocities. These velocities are then categorized into semantic motion classes (halt, slow, normal, fast) using hierarchical clustering. A density-aware, percentile-based scoring system identifies deviations from normal patterns, providing interpretable anomaly detection even with severe occlusions and varying crowd densities.
Detecting unusual events or behaviors in large crowds is a significant challenge, especially in very dense environments where people often block each other from view, and motion patterns can change dramatically based on how packed the area is. Traditional methods often struggle with these complexities, failing to adapt to different crowd densities and lacking clear ways to explain why something is considered anomalous.
Addressing these limitations, researchers have introduced VelocityNet, a novel framework designed for real-time anomaly detection in crowded scenes. This system employs a unique dual-pipeline approach that combines head detection with dense optical flow to precisely measure the velocity of each individual in a crowd. By focusing on head detection, VelocityNet can track individuals more reliably even when their bodies are heavily obscured.
The core of VelocityNet’s approach involves categorizing these person-specific velocities into easily understandable motion classes: ‘halt’, ‘slow’, ‘normal’, and ‘fast’. This is achieved through a process called hierarchical clustering. Following this, a clever percentile-based scoring system is used to identify anomalies by measuring how much an individual’s motion deviates from what is considered ‘normal’ for that specific crowd density. This means the system can adapt its definition of ‘normal’ motion based on whether the crowd is sparse or extremely dense, preventing common movements in a packed area from being mistakenly flagged as unusual.
The framework’s architecture is divided into several key modules. First, the Motion Estimation Module uses a technique called RAPIDFlow to calculate pixel-level movement between video frames, capturing the overall flow of motion. Simultaneously, the Head Detection Module, powered by YOLO11, identifies and localizes individual heads in the scene, even under heavy occlusion. These two streams converge in the Velocity Estimation Module, where the optical flow data is cropped to each detected head’s region, averaged to estimate raw per-person velocity, and then normalized to correct for perspective distortions, ensuring depth-invariant velocity measurements.
Finally, the Anomaly Detection Module takes these normalized velocities. It uses K-means clustering to group similar motion patterns, and then hierarchical clustering to map these groups to the semantic categories (halt, slow, normal, fast). Crucially, it incorporates a density-aware modeling component that trains separate models for low-to-medium and high-density scenes, making the anomaly detection more accurate and robust. The anomaly scoring mechanism then assigns positive scores for unusually fast motion and negative scores for unusually slow motion, relative to the established ‘normal’ range for that crowd density, providing intuitive and interpretable outputs.
The effectiveness of VelocityNet was demonstrated using a unique dataset collected from the Holy Mosque in Makkah, an environment known for its exceptionally dense crowds and constrained pedestrian motion. This real-world testbed allowed the researchers to evaluate the system under challenging conditions, confirming its ability to detect diverse anomalous motion patterns in real-time. The research paper, which provides a detailed explanation of this innovative framework, can be found here.
Also Read:
- Advancing Gait Recognition with Adaptive Motion Analysis
- Time-Conditioned Contraction Matching: A New Era for Scalable and Explainable Anomaly Detection
This work represents a significant step towards robust anomaly detection in densely crowded scenes, offering a solution that is both computationally efficient and provides interpretable results, making it suitable for practical deployments in critical environments.
