A Drone-Person Tracking System in Uniform Crowd: A Deep Learning Vision Approach

Abstract

Robots capable of tracking and following individuals have numerous applications, including autonomous driving, surveillance, and security. However, monitoring people in uniform crowds poses significant challenges that have yet to be adequately addressed. Despite the prevalence of uniform crowds in various contexts, such as the United Arab Emirates (UAE) and Gulf regions, the difficulties they present remain largely unexplored. While ground robots have been utilized in tracking scenarios, most cases involve dispersed populations. Furthermore, the movement of a tracking robot in close proximity to individuals can be disruptive and potentially hazardous. In this context, the use of drones offers a more suitable solution, as they are less prone to collisions, less intrusive, and provide more diverse tracking capabilities compared to vision-based systems. 1) Lightweight Face Recognition: We design a novel family of lightweight FR models, named GhostFaceNets, specifically tailored for efficient computational requirements. These models demonstrate exceptional effectiveness in FR tasks, achieving State-Of-The-Art (SOTA) results on 9 benchmark datasets while minimizing computational requirements, measured by the number of FLoating-point OPerations (FLOPs). By optimizing the balance between accuracy and efficiency, our models prove their efficacy in real-world scenarios. 2) Aerial Face Recognition: Face recognition (FR) in aerial imagery presents unique challenges, notably the impact of Low Resolution (LR) and varying pitch angles. Image quality, particularly resolution, significantly influences the performance of contemporary face recognition systems. However, there has been a limited exploration of how FR models perform on drone-captured images, primarily due to the scarcity of suitable datasets. To address this gap, this study investigates the efficacy of FR models when applied to drone-captured facial images. We employ state-of-the-art (SOTA) lightweight and large FR models and evaluate them across three drone-captured FR benchmarks, in addition to one dataset focused on LR FR. Furthermore, to ensure a comprehensive evaluation, we adopt seven widely recognized benchmarks and artificially downsample and rotate them. This approach simulates the impact of distance and altitude on the view from a drone to a target. The results underscore a substantial decrease in accuracy across all FR models in these scenarios. In response to this challenge, we introduce a novel FR model, EfficientFaceV2S, which consistently demonstrates SOTA performance across all benchmarks while imposing minimal computational demands. This characteristic makes it particularly suitable for real-time and resource-constrained applications. Consequently, this work provides valuable insights for selecting the most appropriate FR architectures for drone-captured images, addressing an emerging need in the field. 3) Drone-Person Tracking Benchmark and a Baseline: To address the unique challenges of drone-person tracking in uniform crowds, we collect a new benchmark dataset, named Drone-Person Tracking in Uniform Crowd (D-PTUAC). This dataset is meticulously designed to simulate intruder behavior within a uniform crowd, considering the specific constraints and dynamics associated with drone-based tracking. It serves as a standardized evaluation platform, enabling researchers and practitioners to compare and evaluate tracking algorithms under realistic conditions. To tackle the problems of multi-scale variations presented heavily in benchmarks that utilizes drones. We propose a pyramidal visual object tracker that leverages progressive shrinking pyramid to reduce the computations, inherits the advantages of both Convolutional Neural Networks (CNNs) and Transformers, which showed usefulness for similarity matching learning.

Date of Award	6 Dec 2023
Original language	American English
Supervisor	Naoufel Werghi (Supervisor)