Target Detection And Trackingbased On Deep Learning

Drone brings convenience to people,at the same time it also brings threat.According to these characteristics of drone,it is imperative to develop corresponding defense system to detect,track and attack invading drone.In this paper,the design of target detection and tracking algorithm in the corresponding system is studied in depth.The main works are as follows:For the problem of long-term target accurate perception in ultra-low altitude target defense system,considering that both target detection and target tracking use convolutional neural network to extract features,and considering with the progress of Siamese network in target tracking field,this paper proposes a fusion algorithm of detection and tracking.Statistical method is used to generate the parameters of candidate regions,and feature pyramid network are used in backbone network.The basic structure of backbone network is inverse residual structure.In the state of target detection,the backbone network is used to extract features from the input image to generate multi-scale feature maps.The target score and position offset are predicted directly on multi-scale feature maps.After the target detected,the weights of the corresponding regions in feature maps are saved.Then the target tracking state is pending.In the state of target tracking,the backbone network is used to extract features from the input image to generate multi-scale feature maps,and then the preserved weights and corresponding feature maps are convoluted.After processing the multi-scale feature maps after convolution,the target score and the regression target position offset are predicted.After the target reliability is lower than the threshold,the target is thought to be lost,and target detection state is entered.Generally speaking,the algorithm combining the advantages of target detection algorithm and target tracking algorithm uses a unified network,similar processes.It has high accuracy,not easy to drift,fast and not easy to miss detection.The skip between the two states is smooth,which meets the needs of ultra-low altitude target defense system.For the problem of data set duplication,the duplication of data sets may cause the internal covariance drift of convolutional neural network,and it is a great waste of time and attention of labelers.This paper proposes a data set duplication removal algorithm based on convolutional auto-encoder,which directly trains convolutional auto-encoder with unlabeled data,and then obtains sparse code of input data by using the part of encoder after training.Convolutional similarity evaluation function is used to evaluate the similarity of different images.In order to obtain accurate and efficient sparse coding,this paper explores the influence of different coders and decoder structures on the algorithm by experimental methods,and then compares the performance of traditional methods.The experimental results show that the redundancy of data sets can be effectively removed based on convolutional auto-encoder and convolutional similarity function.As for the difficult problem of data annotation acquisition,the author of this paper finds that the efficiency of modifying annotation is higher than that of original image annotation from scratch.Inspired by this,this paper proposes a progressive annotation strategy.Firstly,the raw data set is divided into ten parts.Firstly,the first part of data is annotated.Then,a model is trained with this part of data set,and the second part of data is predicted with this model.According to this,modify the labeling of the second data.Finally use the first two part of data to train the second weak model.And so on,until all data are modified.Practice proves that the progressive annotation strategy is effective.For the problem of poor detection accuracy of small targets,this paper finds that the main reason for the poor detection accuracy of small targets is that it is unfair for small targets to use Io U as a measure criterion when judging positive and negative samples.Even if the small targets are completely in the candidate area,the Io U between the small targets and the candidate area is very small,it is still considered as negative samples,resulting in poor detection accuracy of small targets.In order to improve the training opportunities for small samples,a new simple and efficient target amplification method is proposed in this paper.A special strategy is to replicate the target randomly,replicate the target at the same time by rotating and adding.In order to overcome the edge effect,the cosine window is used to add the background to the edge when pasting.Experiments show that the detection accuracy of small targets improved by the new augmentation method.