Detecting Social Groups Using Low Mounted Camera in Mass Religious Gatherings

People often come in groups rather than individuals in public places. Automatic visual group detection could help in understanding the group patterns in such scenarios. Researchers have proposed various approaches to identify group patterns based on trajectories extracted from video. However, these pixel trajectories are used without considering perspective projection which results in erroneous data. Placing the camera at a higher point resulting in a top-down or bird’s eye view is often used as an alternative method to circumvent this problem. However, this may not be possible in all real-life scenarios. Thus, we propose an approach for developing simple group detection models based on real-world trajectories by applying projection transformation on the pixel-based trajectories, which can be applied to the data collected using low-mounted cameras. It was found that parameters like average distance between two pedestrians, average angle between their direction of motion for a specified time window is more evenly distributed while using pixel-based location rather than their real counterparts. It is observed that the model performance is improved with the increased time window. Also, simplistic models with real-world trajectories were able to achieve over 90% accuracy.