Research On Key Technologies Of Real-time Multiple Object Tracking Based On Deep Learning

Real-time tracking of multiple moving objects is a key issue in the field of digital image processing,and has broad application prospects in military,industrial and other fields.The feature expression of the target to be tracked in the field of view directly affects the performance of multi-object tracking.However,the problem of occlusion between the foreground and background and occlusion between multiple targets makes it difficult to extract the features.Deep learning is a complex machine learning algorithm.It learns the inherent rules of sample data to obtain the deep-level features of the sample,and then better completes the interpretation of the sample.In recent years,thanks to deep learning mining of deep features of images,the accuracy of image processing technology based on deep learning has been greatly improved,but mostly used in areas such as target detection or classification and rarely used in multi-object tracking tasks.However,as an image sequence information processing technology,multi-object tracking has higher requirements for real-time performance.Therefore,how to use the advantages of deep learning for image information mining and meet its requirements for real-time performance,while improving the accuracy and speed of the algorithm,there exists huge challenges.To that end,research on real-time multi-object tracking technology based on deep learning has strong academic significance and practical value.Based on the analysis of the key technologies of multi-object tracking based on deep learning,this paper studies the key technologies of enhanced multi-object tracking based on single object tracking,global optimization algorithm,motion model,and generative model for data generation.This thesis mainly completed the following four aspects:1.In-depth study of the single object tracking method for single object context information processing and foreground / background differentiation,analyzes the problem that the current multi-object tracking method based on the detection-beforetrack strategy is too dependent on the detector results,an enhanced multi-object tracking algorithm based on single object tracking is proposed performed on the MOT Challenge 17 public data set.Quantitative and qualitative analysis is performed.Experiments show that the method has a MOTA of 48.9 and an IDF1 of 57.0,both of which can reach the state-of-the-art algorithm level.Among them,IDF1 can surpass offline algorithms and achieve the best results of all algorithms.Compared with the algorithm with similar numbers of false detections and missed detections(ie,the environment is similar),the method has improved the IDF1 of the object identification accuracy and the MT indicator of the object identification stability by 22.2% and 17% respectively.It is proved that the method can improve the accuracy of object identification and maintain the stability of object identification while maintaining tracking accuracy.2.In-depth study of the candidate box prediction method based on regression,analyzed the current multi-object tracking method's insufficient use of global information,and proposed a global optimization based on the combination of deep architecture regression model and classifier based on Siamese Neural Networks.The algorithm is verified on the pedestrian re-identification datasets DukeMTMT-reID,Market-1501 and CUHK03 datasets,and the MOT Challenge dataset.This method can make full use of the global information in a single frame,avoid the impact of the uncertainty of the number of objects on the complexity of the tracking algorithm;and use a Siamese Neural Network based on single-sample training to avoid the impact of the update process and low-confidence detection information pairs.This method has a good effect in maintaining long-term tracking status of objects,dealing with occlusion problems,and object interactions.The specific object identification accuracy(IDF1 is 57.3)reaches the most advanced level,and the multi-object tracking accuracy(MOTA is 55.4),The maximum tracking trajectory(MT is 21.3%)is 3.6% and 9.2% higher than the optimal online algorithm,and the object identification exchange(ID Sw.Is 1205)is reduced by 27.9%.3.In-depth study of motion model based on recurrent neural network,analysis of the shortcomings of object motion estimation in the multi-frame environment of the commonly used constant speed model,and an object motion model based on recurrent neural network is proposed.The method is experimentally verified on the MOT Challenge 17 training set and test set for object movement trend prediction and replacement update strategy.By integrating this method into the multi-object tracking algorithms proposed in Chapters 2 and 3 of this paper,the IDF1 indicators representing long-term tracking performance and processing occlusion ability are increased by 2.5 % and 3% respectively,and the ID Sw.is reduced by 15.9%;and after replacing with the multi-frame update strategy,the two multi-object tracking algorithms integrated with the motion trend prediction model can increase the running speed by 229% and 400% respectively without affecting the tracking performance.4.In-depth study of generative methods and analysis of sample characteristics in existing data sets.Focus on the diversity and lack of environmental samples for multiobject tracking,a generative model based on conditional variational autoencoderconditional generative adversarial network is proposed,and the performance of the generative model is verified using datasets targeting pedestrian re-identification task.Experimental results show that this method can take advantage of both models to improve the quality of the results.Next,the proposed generative model is applied to the two tracking methods proposed in the paper: a method based on correlation filters and a method based on Siamese Neural Networks.For the online update method,the model is used as an important means to enrich the training space;as for the offline training method,the model can be used as a method for training data generation to optimize the tracking algorithm.Finally,the specific generative method is introduced into the multiobject tracking methods proposed in Chapters 2 and 3 of this paper for performance verification.After using this method to expand the diversity of data,it can effectively reduce the possible deformation and the effect of occlusion.Statistics show that it can specifically reduce false negatives(FN)by 3.1% and 4.8%,and reduce Identity Switch(ID Sw.)by 22.5% and 18.3% respectively.This method can effectively enrich the diversity of samples in the object tracking dataset,reduces the impact of complex environments on algorithm performance and improves the overall tracking performance of the multi-object tracking method.This thesis addresses the shortcomings in the research of key technologies for multi-object tracking at the current stage,from improving the long-term tracking stability,the tracking accuracy through global optimization,he accuracy of the motion model to reduce the calculation cost of the algorithm,and increasing the diversity of data samples to ensure that the algorithm in terms of stickiness.A variety of algorithms are proposed to improve the performance of multi-object tracking methods.At the same time,a large number of simulation experiments were performed on each proposed algorithm.Experimental results show that the methods proposed in this paper can effectively improve the tracking performance of the algorithm for various key technologies in multi-object tracking tasks.