Research On UAV Target Detection Algorithm Based On Deep Learning

UAV,as a new tool of information and material transmission,has developed rapidly in recent years.With the development of UAV application technology,it is possible to use UAV to carry camera for aerial photographing,security line inspection,circuit inspection and search and rescue.As an important part of UAV application,UAV target detection has been a hot topic in scientific research.Considering the constraints of the calculation,the power loss and working environment of the UAV airborne processor,the target detection algorithm embedded in UAV needs to meet the requirements of less parameters,less memory consumption and high real-time performance.In view of the working environment of UAV and the large change of object scale,the algorithm of UAV needs to ensure the performance of small target detection while conventional target detection.Therefore,how to simplify the algorithm model and to consider the performance of target detection becomes an urgent problem.Based on the Yolov4-tiny target detection algorithm and the thought of "think twice",this paper proposes an improved RFPYolov4-tiny algorithm based on recursive features.This algorithm can meet the needs of lightweight and less parameters,and also has good detection accuracy,and can enhance the accuracy of small target detection.The main contents of this paper are as follows:(1)Starting from the target detection algorithm based on deep learning,this paper deeply analyzes the model structure of various versions of the algorithm,compares the advantages and disadvantages of the models,and selects the lightweight model yolov4 tiny,which is more suitable for embedded UAV and other mobile terminals,as the algorithm model.The experimental results show that as an algorithm embedded in UAV equipment,yolov4 tiny has more value,and provides a good research foundation for the subsequent algorithm improvement.(2)In view of the accuracy of target detection and small target detection,rfpyolov4 tin algorithm is proposed in this paper.The algorithm improves the backbone extraction network and FPN of yolov4 tin algorithm.The main network is modified as a circular recursive network,so as to realize the feature extraction twice and strengthen the semantic expression.At the same time,ASPP is added as feature conversion,and feature fusion module is introduced to form the network structure of rfpyolov4 tin algorithm.The improved method proposed in this paper,the detection time of single image is 26 ms,and the algorithm has good performance in the situation of occlusion and small target detection.(3)For the lightweight of target detection algorithm,this paper proposes rfpyolov4 tin optimization strategy based on deep separable convolution,and replaces the ordinary convolution in the algorithm with depth separable convolution,so as to simplify the model.Compared with rfpyolov4 tin,the lightweight model has a good improvement in parameter and operation speed,which provides a reference for the improvement of the target detection algorithm of UAV.