Convolutional Neural Network-Based Multi-Target Detection and Recognition Method for Unmanned Airborne Surveillance Systems

Abstract

This paper proposes the convolutional neural network (CNN)-based multiple targets detection and recognition method for unmanned airborne surveillance systems. The proposed method is capable of recognizing the target's type, position and bearing angle. Recently, deep learning approaches using convolutional neural networks (CNNs) have significantly improved the object detection accuracy on benchmark datasets such as Pascal visual object classes (VOC) and common objects in context (COCO) data sets. Typical CNN-based object detection technologies are designed to recognize regions of interest (RoI) and object classes based on VOC or COCO data set criteria only. However, in many surveillance missions, the bearing angle of the object is also an important entity to infer in addition to the RoI and the vehicle-type. This paper proposes a CNN-based object recognition technique called airborne surveillance neural network (ASNet) that can recognize this additional bearing angle information. Indoor experiments demonstrate the validity of the proposed method.