Person Localization and Crowd Counting using Deep Learning (Aerial images)

• 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.

Proposed Platform for Person Localization and Crowd Sourcing

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

Results

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