Research On Particle Filter Target Tracking Algorithm Based On Deep Learning

Target tracking,as a research hotspot in the field of computer vision,has a very wide range of application prospects in practical scenarios such as intelligent surveillance,human-computer interaction,and military reconnaissance.Target tracking is affected by scale changes,illumination changes,occlusion,etc.,and deep learning Based on the network framework,single-target tracking is studied from the aspects of feature extraction and model update.The main research contents are as follows:(1)Aiming at the problem of insufficient training samples in the target tracking process and the poor classification effect of traditional classifiers in the case of linear inseparability,a particle filter target tracking algorithm based on deep noise reduction autoencoder is proposed,which uses stack noise reduction Autoencoder(Stack Denoising Autoencoder,SDAE)as a deep network model,through the noise processing of the data,offline training on the neural network model,so as to improve the network's characterization ability and anti-interference ability.Support Vector Machine(SVM)and radial basis kernel function are used to realize the classification of features and improve the classification ability of the network and the accuracy of the tracking algorithm.In the tracking process,the particle filter is used to disperse the particles,and the confidence of the particles is calculated through the deep network model to achieve effective tracking of a single target object.(2)In order to solve the problem of algorithm accuracy and high computational cost,an improved particle filter target tracking algorithm based on convolutional neural network is proposed.First,the attention mechanism(Attention Model)is used to constrain the sampled particles.By adding an attention mechanism layer after the feature map,the interference features are suppressed,and the feature extraction ability of the convolutional neural network is improved.The training efficiency and generalization ability of the network.Secondly,the simple and efficient Softmax is used as the classifier to improve the calculation speed and classification ability of the network.Finally,the accurate tracking of a single target object is realized under the framework of the particle filter algorithm.At the same time,the model is updated by long and short-term decision-making.The training method of sample mining(Online Hard Example Mining,OHEM)reduces the computational cost of the algorithm and improves the ability to distinguish difficult samples.(3)Comparing the OTB50 data set with other mainstream tracking algorithms,the experimental results show that the two deep learning-based particle filter target tracking algorithms proposed in this paper can be implemented under the influence of external factors such as scale changes,illumination changes,and occlusion.For the accurate tracking of the target,the accuracy rate increased by 4.48% and 4.86% respectively,reaching 69.1% and 82.7%.Effectively improve the robustness of the target tracking algorithm.