- Person localization and crowd sourcing in aerial images as many applications including crowd safety and security management, people management in disasters, and person movement in an epidemic disease situation.
- Ariel images have an issue of cluttered background, low-resolution quality, and various environment and lighting conditions, which increase complexity levels of localization and recognition.
- Deep learning and a UAV Raspberry Pi-based platform for the acquisition and localization of crowds considering the different orientations of the UAV and heterogeneous environmental factors.
- The proposed system consists of: the transfer of the image from the UAV to the processing system, image enhancement, features extraction for object detection and recognition, and performance comparison of different detection and recognition algorithms.
- We collect a large local environment dataset for better accuracy.
- We compare the performance of Haar-Cascade, template matching, YOLO, and Mask-Region Based Convolutional Neural Networks.
- The aforementioned results show that YOLO performs better for crowd detection and localization.
Fig. 1: Proposed Architecture of Deep Learning Platform for Person Localization and Crowd Management in UAV Aerial Images
Fig. 2: GUI (Graphic User Interface) of Person Localization and Crowd Sourcing in Aerial Images
Algorithms | No of Person (Detected) |
---|---|
Haar Cascade | 1530 |
You Only Look Once (YOLO) | 2524 |
Mask RCNN Algorithm | 2495 |
Fig. 3: Results of Person Detection and Crowd Sourcing
Fig. 4: Graph of actual persons, and haar-cascade, YOLO, mask RCNN algorithm detected persons