Deep Neural Network Based Multi-Object Detection for Real-time Aerial Surveillance

Abstract

Aerial surveillance is one of the widely used modern days surveillance methodologies, finding applications in many important fields including military and civilian. This article presents a comprehensive study of Deep Neural Network (DNN) based solutions for real-time object tracking from Unmanned Aerial Vehicle (UAV) using a modified version of the state-of-the-art object detection algorithm YOLOv5 model. The modified YOLOv5 architecture is achieved by changing the activation function to Rectified Linear Unit (ReLU) and fine-tuning the network’s hyperparameter. A comparative analysis was then done on a subset of the AU-AIR dataset by comparing the different YOLOv5 models based on the network depth to determine the improvements in training speed and accuracy. The modified network was also compared in terms of mean average precision (mAP) to the original paper, a performance gain of almost 2.9 times was achieved in the best-case scenario.