An Airborne Multi-spectral Target Detection System Based On Deep Learning

In recent years,the Unmanned Aerial Vehicles(UAVs)have developed rapidly in civil and military fields.The UAVs equipped with a variety of sensors can achieve all-weather perception of ground scene information,and autonomous flight of UAVs can be realized through a series of decision algorithms.An airborne multi-spectral target detection system based on deep learning is designed and developed in this paper.The system acquires image data through infrared detectors and visible light cameras mounted on the drone platform,and performs real-time reasoning on the embedded platform,then the results of target detection are fed back to the flight control module,so as to control the flight direction and distance of the aircraft and realize the autonomous flight of UAVs.The image processing subsystem is an important part of the system,and its core is the target detection algorithm for aerial images.Due to the great change for the scene of aerial image and the lack of detail information,it is difficult to extract the target effectively by traditional methods,and the detection framework based on deep learning can extract features better.However,due to the limited computing resources of the embedded platform,it is difficult to meet the needs of deep network fast reasoning.On account of the characteristics of aerial photography data and the limitation of resources on the embedded platform,an airborne multi-spectral target detection algorithm called CENet-SSD based on deep learning is proposed in this paper.This paper combines the ideas of DenseNet and ResNet to propose a low-parameter named CE module to construct a feature extraction network to improve network performance and efficiency.At the same time,redundant preselected boxes in the prediction of layer are removed to effectively reduces the amount of calculation.And the experiment result shows that this method not only ensures the accuracy,but also greatly improves the speed of the algorithm.Besides,in order to further improve the efficiency of the algorithm,the network reasoning acceleration scheme based on TensorRT is also used in this system,which accelerates CENet-SSD network model by using highly integrated kernel modules and reducing weight precision,finally implements the high speed reasoning of 60 FPS on the embedded platform Nvidia Jetson TX2.