UAV Detection Method Using Convolution Neural Networks Optimized By Hyper Parameters Tuning

Abstract

Listen

Translate article

Recently, a great advantageous of autonomous unmanned air vehicles (UAVs) have increased widely due to device low cost and smaller size. Unmanned aerial vehicles (UAV) are commonly exploited to provide abundant helpfully services such as package delivery and surveillance systems. However, the possible threat to public security and personals live occasioned by drones intrusion should be addressed as a critical issue needs to resolve. In this paper, a new deep learning model for UAV detection using a customized convolution neural network CNN named IC-CNN is designed and introduced. The main objective of this research is to detect the drone object appearing at variant locations in the acquired images with different resolutions. Meanwhile, birds’ objects may possibly present in the image foreground. An optimal configuration of hyper-parameters adopted in this work is conducted to get higher accuracy of drones’ detection. The proposed deep learning model adopts forward and backward training workflow with customized layers for detecting drone object in the acquired images and distinguishing drone object from bird object. An optimization process is achieved to select accurate hyper-parameters that effected mainly in the classification accuracy. IC-CNN deep learning model is constructed by stacking sequence of convolution; maximum pooling and fully connected layers integrated in light weight network architecture. The proposed deep learning model based UAV detection system is evaluated using three types of drones’ image datasets along with common evaluation metrics. The obtained results achieved 88%, 100% and 100% detection accuracy on the three datasets respectively.