Research On Tracking Algorithm Of Dim Target Baserd On Deep Learning

Dim target tracking plays an important role in radar detection,machine vision,medicine and other fields.However,in practical application,due to the influence of background clutter,low signal-to-noise ratio,occlusion deformation and other factors,dim targets are easy to be interfered by strong noise in the surrounding environment.At this time,the algorithm which is too simple to deal with a single scene or scene is difficult to apply.At present,the problem of target tracking based on multiple application scenarios has become the focus of research.The scenarios are mainly divided into three categories: sky model with cloud interference,sea sky model with sea interference,and land air model with ground object interference.In order to solve the problem of interference caused by complex scenes on weak target tracking and improve the tracking accuracy and accuracy,on this basis,the dim target tracking algorithm based on deep learning is studied1)In the process of image preprocessing,this paper proposes a maximum variance threshold segmentation algorithm based on background estimation.The experimental results show that the improved algorithm can better obtain the feature information of candidate areas;the improved guided image filtering algorithm,through experimental comparison,shows that the improved algorithm can obviously highlight the information of dim targets in the image and improve the detection accuracy.2)The cloud interference suppression subsystem is designed.In order to solve the problems of less feature information of dim target and low efficiency of training sample extraction in sky model,the guide image filter is improved to filter the target image,which makes the inaccurate background template fuzzy and enhances the target image effectively.The improved denoising autoencoder overcomes the shortage of training sample.Compared with many mainstream algorithms,the system can effectively suppress the interference factors such as the gray fluctuation in the cloud,strong gray noise and edge cloud occlusion,and is more suitable for the appearance change of the target in the tracking process.3)The sea interference suppression subsystem is designed.In view of the interference of the sea sky boundary to the dim target,the feature extracted by the maximum variance threshold segmentation algorithm of fusion background estimation is used as the input of the depth autoencoder for training.The experimental results show that the algorithm can suppress the influence of long and thin interfering targets.Compared with the first-order and second-order differential edge operators,the system has higher accuracy in extracting target candidate regions.4)The interference suppression subsystem is designed.In view of the interference of complex targets in the near scene,the RGB-depth image feature is obtained by convolution autoencoder as the high-level feature value.Then,according to the characteristics of weak target image,the 12-dimensional low-level feature vector is selected,and the skip layer structure is added in the network training process to fuse the multi-scale depth features.Experimental results show that RGB-depth image features can effectively eliminate the interference of complex targets in close range.5)Three algorithms are integrated into a high-performance target tracking algorithm.Soft Max classifier responds to the corresponding subsystem according to the model category,effectively avoiding the impact of different environments on dim target tracking.The trained deep learning network and classifier are applied to the particle filter framework to form a dim target detection system based on deep learning.In order to verify the feasibility of the system,several methods such as overlapping rate chart,accuracy chart,success rate curve,time and space robustness evaluation are used to analyze the algorithm.The experimental results show that the integrated algorithm has better tracking accuracy in the process of dim target tracking.