Study On UAV Target Tracking Algorithm Based On Artificial Neural Network

Target tracking is currently widely used in areas such as unmanned driving,video surveillance,and human-computer interaction.Since the field of target detection is relatively mature,experts and scholars at home and abroad have paid more attention to the field of target tracking.In these fields,target tracking first determines the position and size of the target,and then locates the target position correctly in a series of frames.With the further landing of driverless technology,future object tracking will still be a very popular research direction,uav target tracking is of great application prospect.However uav target tracking with occlusions,small target,illumination change,scale changes,and the resource limitation in the uav platforms,aiming at the uav application scenario,this paper makes a series of improvement research on the platform resource limitation and target small challenges in uav target tracking.In the past,the methods based on correlation filter performed well in the target tracking benchmark test,but most of these methods only used artificial features to represent the tracking object,which limited the representation ability of the object and often failed to achieve satisfactory performance under the interference of factors such as occlusion and background clutter.In recent years,the target tracking algorithm based on siamese network began to appear in the public view.Among them,the most representative one is based on full convolution siamese network tracking algorithm(SiamFC),and then developed a series of twin network tracking algorithm.SiamRPN++ introduces the deeper benchmark network ResNet into the target tracking,which is currently a tracking algorithm with strong comprehensive performance.However,with the deepening of the network,the greater the consumption of computing,it can not meet the use on the drone platform;At the same time,due to the small target in the uav scene,the characteristics of different layers need to be used more efficiently.Therefore,in this paper,based on the SiamRPN++ framework,the existing problems in the uav scene are optimized by two methods:(1)Compression of the feature extraction network ResNet using geometric median method,so that the model runs in the drone scene,while controlling the tracking accuracy within an acceptable range;(2)Research on target tracking of siamese networks based on deep feature fusion weighting.In view of the shortcomings of SiamRPN++ in multi-layer feature fusion,an improved deep feature fusion weighting method is introduced.Using the convolutional neural network training weights makes the algorithm more accurate for the classification and regression results.Finally,in order to verify the effectiveness of the algorithm,an improved siamese tracking framework is established and experiments are performed on the data set.The results show that compared with other original algorithms,the model is smaller and more robust,and the accuracy is acceptable without reducing the tracking rate.There are 20 figures,6 tables and 78 references in this paper.