Optimization Method Of Ground Target Detection For Micro-UAV

In recent years,with the rapid development of micro UAV technology,the application of micro UAV has achieved fruitful research results in civil and military fields,and it has broad prospect.The onboard visual perception system of UAV is the integrated system of "eyes" and "brain" which can directly caputure the environment image and extract the high-level effective information,this information can lead other UAV missions such as battlefield investigation.Onboard visual perception system is the necessary part for the micro UAV to complete various tasks,and it is also the comprehensive embodiment intelligence level of UAV.Micro UAVs have weak load capacity and portable resources are very limited.This requires higher integration of onboard visual sensing system.For the application requirements of accurate detection,recognition and location of ground targets in field environment,this paper constructs a low-power hardware system under the limit of extremely weak load capacity of micro UAVs.A lightweight ground object detection and recognition algorithm is designed for the onboard hardware system,and researches are pursued around the detection and recognition algorithm with the hardware system.Visible light camera is selected as the main sensor to capture environmental information,and a lightweight embedded airborne processor is used as the computing platform.Machine learning,filtering,optimization and other technologies are used to detect the accurate position of the target in the image in real time,which is used for the final target space position estimation.The main work and innovations of this paper are summarized as follows:(1)constructing lightweight hardware and software architecture for onboard visual sensing systemThere are some defects in load,power consumption and volume of micro UAV,which restrict the hardware and software of the sensing system.In this paper,a lightweight onboard visual sensing software and hardware system is designed for micro UAV.In terms of hardware architecture,the low-power embedded processor is used as the processing center of the onboard visual sensing system,and the airborne visible light camera with light weight,small size is used to perceive the surrounding environment.In addition,in order to enhance the computing ability of the system without too much increase in load and power consumption,neural computing chip is introduced to expand the hardware architecture of the sensing system,which greatly improves the computing speed of the neural network.In terms of software architecture,a lightweight detection-tracking-localization algorithm framework is designed.The saliency region proposal method and the shallow convolution neural network are used to accurately detect and identify the target in real-time,and the filtering and optimization methods are combined to achieve the accurate spatial positioning of the target.In short,the whole perceptual hardware and software system is light in weight,small in volume,low in power consumption and low in computing resources,which fully meets the application requirements of micro UAV.(2)Sample optimization method for airborne target recognition networkThe imbalance of training data is easy to occur in flight experiment,which will affect the training effect of convolutional neural network and bring performance loss to the perceptual system.To solve the problem of unbalanced training data,this paper optimizes the training sample structure of convolutional neural network,and improves the target recognition accuracy of convolutional neural network from the sample structure level.Sample optimization method can maximize the performance of the airborne target detection and recognition algorithm without increasing the amount of computation and the load of the airborne processor.According to the difference of data status in flight experiment,two ideas are designed to optimize the sample.In the case of poor flight data content,the quantity balance strategy is used,including selecting video samples according to the time and optimizing the samples through the quantity balance processing on the target class.When the content of flight data is gradually rich,the method of process balance is used to optimize the samples,including selecting the flight video content through the method of factor combination.By analyzing the distribution of the target at different illumination angles,the angle balance of the unmanned aerial vehicle(UAV)hovering flight data is processed.The optimized training samples obtained by the two strategies can get better network weights in training and these weights can improve the accuracy of the airborne detection and recognition algorithm.(3)onboard target recognition convolutional neural network optimization methodConsidering the limited improvement for object recognition accuracy using the proposed sample optimization scheme,the recognition accuracy could be further improved by increasing the number of convolutional neural network layers and optimizing the convolutional kernel distribution on the extended hardware architecture(Network Process Units).Firstly,the CaffeNet was introduced,which greatly promote the recognition accuracy.However,it did not meet the real-time capability requirements even running on the extended hardware architecture.To reduce the computational complexity of the network without significantly decreasing the recognition accuracy,the input size of the network and the convolutional kernel distribution were refined,according to the size and imaging characteristics of the objects in the onboard images.These operations significantly speeded up the network running,so that its real-time capability fully satisfied the requirements of the subsequent localization module.