Research On Recognition Method Of Engineering Vehicle In UAV Aerial Image

When the engineering vehicle is working under or near the high-voltage power line,the distance between its bucket or boom and the high-voltage power line would probably to be less than the safe distance,which is very easy to touch the power line and result in safety accidents such as short-circuit breakdowns.Therefore,it is essential to detect the engineering vehicles that work near the high-voltage power line.Unmanned Aerial Vehicle（UAV）is currently widely used in power inspection because of its high flexibility and high inspection efficiency.In the inspection process,targets are detected through the recognition algorithm deployed on the UAV’s airborne platform.However,many traditional target recognition algorithms may meet many problems while they are used to recognize the engineering vehicle from aerial image such as inaccuracy and inefficiency.Therefore,it is necessary to explore an algorithm with high accuracy and high efficiency for the recognition of engineering vehicles in UAV aerial image.For this purpose,the capsule network algorithm is studied in this thesis,and the network structure and dynamic routing algorithm of the capsule network are improved in view of the characteristics of small parameter scale and high target recognition accuracy with small dataset.At the same time,it is compared with the classical pattern recognition algorithm and YOLOv5（You Only Look Once Version 5）algorithm respectively,in order to explore an optimal capsule network model for recognizing engineering vehicles in aerial images.This research was supported by the Guangzhou Science and Technology Project（No.202007040004）.The main contributions of this thesis are as follows:1.Design the overall architecture of engineering vehicle recognition in UAV aerial images.2.Constructed UAV aerial images dataset of engineering vehicles.Firstly,the data was enhanced by image preprocessing operations such as adding noise and adjusting brightness.Secondly,the images were annotated.This thesis built two datasets,of which dataset 1 was composed of annotated images,and dataset 2 was the dataset obtained after cropping and normalizing the annotated images.Two datasets were grouped into training set and test set respectively.3.Constructed the capsule network algorithm model for recognizing engineering vehicles in UAV aerial image.Firstly,the network structure of capsule network is improved,that is,the original capsule network structure is improved to multi-layer densely connected capsule network,referred to as the Improvement 1 algorithm.Secondly,the dynamic routing algorithm of the capsule network is improved,that is,the Soft Max function in the dynamic routing algorithm is replaced by leaky-Soft Max function,referred to as the Improvement 2 algorithm.Then,the Improvement 1 algorithm is combined with the Improvement 2 algorithm,referred to as the Improvement 3 algorithm.4.The key parameters of capsule network,such as network layers,routing coefficient and squeezing coefficient,were studied to explore the optimal algorithm model of capsule network,and compared with the classical pattern recognition algorithm model,YOLOv5 x model and original capsule network model.5.Considering that capsule network needs to be combined with target region generation method in target detection,however,traditional target region generation method is usually time-consuming in the process of target region generation.In order to improve the real-time performance of capsule network target detection,RPN（Region Proposal Network）was combined with capsule network to build a whole-process target detection model of engineering vehicle recognition in UAV aerial images,and its target detection performance was compared with the capsule network target recognition model combined with SS（Selective Search）.The experimental results show that:（1）The improved methods proposed in this thesis can effectively improve the target recognition accuracy of the capsule network,all three improvements make the mean Average Precision（m AP）of the capsule network higher than 90%,wherein the m AP of the Improvement 1 algorithm is 91.70%,the m AP of the Improvement2 algorithm is 90.03%,and the m AP of the Improvement 3 algorithm is 92.10%,which id2.61% higher than the original capsule network.（2）Different capsule network layers will lead to different image feature extraction effects.In this thesis,the target recognition performance of the capsule network model with 1,3,5 and 7 layers is compared respectively.The results show that the number of layers of capsule network has a great influence on the recognition accuracy,but the relationship between them is non-monotone and nonlinear.In the application scenario of this thesis,when the number of network layers is 5,the m AP of capsule network is highest.In addition,the parameter scale will not be increased by using the multi-layer densely connected capsule network.（3）The routing coefficient and squeezing coefficient of the dynamic routing algorithm of capsule network would affect the recognition accuracy of capsule network.After exploration,when the routing coefficient is 5 and the squeezing coefficient is 1,the Leaky-Soft Max 5-layer densely connected capsule network has the highest target recognition accuracy,with m AP reaching 94.56%,which is named as the optimal algorithm model for the capsule network.（4）The optimal algorithm model in this thesis is obviously better than the classical pattern recognition algorithm and YOLOv5 x algorithm in m AP performance,and the parameter scale is less than one third of YOLOv5 x algorithm.（5）The whole process detection algorithm model of capsule network combined with RPN constructed in this thesis can detect targets in an image with a resolution of 2592× 1944 Pixels every 2.55 s on average,while the capsule network combined with SS can detect a target in an image with the same resolution every 44.70 s on average because its process of generating candidate regions is time-consuming.Therefore,the whole process detection model of capsule network combined with RPN constructed in this thesis is obviously superior to the two-step detection model of capsule network combined with SS in real-time performance.In summary,in the case that the UAV aerial image dataset containing engineering vehicles is small,this thesis improved the capsule network and explored the key parameters of the capsule network,and proposed an optimal algorithm model of the capsule network.The parameter scale of this model is small,and the target recognition accuracy is obviously better than other algorithms in the application scenarios of this thesis.Meanwhile,its real-time performance can basically meet the engineering needs,so it can realize the accurate and efficient detection of engineering vehicles in UAV aerial images.