Research On UAV Target Detection Algorithm Based On Convolution Neural Network

Target detection is to obtain the image or video sequence of the information such as geometric features and statistical characteristics of a particular target with digital image correlation technique,realize the goal of segmentation and extraction,and then complete the follow-up process of recognition,tracking and other tasks.With the current unmanned aerial vehicle technology constantly making breakthrough progress,the domestic unmanned aerial vehicle(UAV)market will also have explosive growth.For the National military applications,UAV have become a new type of air support force,but target detection is also key technology.Also along with the rapid development of computer science technology and digital image processing technology which are widely used in various fields,using digital image processing technology for UAV for the interested target in the image in real-time and efficient detection also gradually become a hot direction of many universities and research institutions.Dynamic,real-time and stable detection and location of specific targets have wide application value in modern traffic management,community monitoring,campus security,national military application and medical auxiliary navigation surgery.However,it is obvious that the intelligent processing capability of UAV's airborne system is insufficient.For example: real-time dynamic target detection,tracking and positioning.Therefore,it is of great theoretical significance and engineering application value to study the high-efficiency and high-precision UAV's target detection algorithm.To solve the above problems based on pedestrian target occlusion processing,deformation and small target detection,this paper proposes a new solution.This solution is based on the convolution neural network target detection algorithm based on the combination of preprocessing and target priority location;the high performance processor of NVIDIA's Jetson TX1 combined with GPU acceleration ensures the smooth and real-time operation of the UAV target detection algorithm which is based on the convolution neural network.First of all,starting from the basic theory of target detection,this paper analyzes the feature extraction in the current detection algorithm,the current situation of occlusion processing,and summarizes the problems and difficulties that are difficult to solve.In the face of the above problems,This paper uses(CNN)convolutional neural network as the basic research tool,and proposes the UAV target detection algorithm and its system implementation based on the convolution neural network.To this end,the research work of this paper mainly from the following two aspects:(1)This paper studies the theory and method of Convolution Neural Network,and improves the original algorithm which needs a large number of data sets.Starting from the feature region of the whole picture,this paper adopts the end-to-end learning algorithm,combining with the neural network architecture of YOLO,set up the improved based on pedestrian detection of network layer,and improve the accuracy and efficiency of target detection.Aimed at the early stage of the input image's pretreatment process,the collected image color space conversion and normalized processing,using(Histogram of Oriented Gradient,HOG)to get a feature map of pedestrian,finally using BING algorithm to filter out image background information,the final results input into the binary image convolution layer,only the characteristic information of the detection target and the contour of some background images are reserved.The network layer draw lessons from the anchor thought of Fater R-CNN,and makes the prediction of the preselection box directly on the output characteristic chart.Use the anchor boxes to predict the bounding box.This paper mainly adopts k-means clustering automatically to obtain the preselected box.Output after the detection of parts with the improved method to predict frame's confidence and Io U,when there is more primary box intersect,the value which has lower confidence will not be deleted from the overlapping box directly.The optimal threshold of the pedestrian target was obtained through the preliminary training,and compare it again to determine whether to retain.Experimental results show that the algorithm has excellent performance in small targets and small amount of feature data of occlusion targets.(2)To test and verify the accuracy and stability of the algorithm proposed in this paper,we compare the proposed model with a variety of existing mainstream target detection methods,such as HOG+SVM,haar-like+Adaboost and YOLO.This paper will verify and analyze the anti-interference,anti-occlusion,stability,real-time and accuracy through the simulation verification experiment,and give the analysis results.Finally,offline training on PC increased weight,and the overall test frame was migrated to NVIDIA's Jetson TX1 processor for system testing and real-time remote interaction with the UAV The processor will be used as a computing processing center for UAV,making the UAV completely free from the limitation of lack of computing power.Thus,an automatic,efficient and fast pedestrian detection system can be used in real environment.