Research On Perception Mechanism And Detection Algorithms For Video Targets Based On Electroencephalogram

With its intuitive and rich information presentation form,video has become an important data source for capturing vital and sensitive information.Important object detection and abnormal behavior analysis in video have a wide application demands in the fields of social security and national defense.However,due to the dynamics of video targets and the complexity of video background,it is difficult to build an effective detection model.Therefore,there is an urgent need to develop more accurate and generalized video target detection technology to provide strong support for its important applications in the fields of social and national security.Recent research of brain science shows that the human brain with long-term evolution will be induced specific neural information by sensitive targets.Based on this neural information,a new “brain-computer interface”(BCI)technology,integrating the advantages of machine intelligence and human intelligence,combining the powerful computing power of the machine and the intelligent perception and logical reasoning ability of the brain in the external complex environment,will bring a new view to break through the bottleneck problem of video target detection,which has important theoretical and application value.Aiming at developing the BCI-based video target detection technology,this study focuses on perception mechanism and detection algorithm for video targets.In terms of perception mechanism,in the cortical source space,using the time-varying brain network analysis method,this study explores the dynamic modes of brain information interaction in the perception process of task targets and non-task targets,which provides an important theoretical support for revealing the video target perception mechanism in complex environment and constructing an effective and reliable detection model;In terms of detection algorithm,aiming at improving the timeliness,stability,and accuracy of video target detection,this study constructs the asynchronous detection model based on single-trial EEG signals,the cross-domain detection model based on EEG signals,and the multi-brain enhancement model based on a mutual learning domain adaptation network,which effectively improve the performance of the BCI-based video target detection.The main works of this study are as follows:1.Research on the perception mechanism of task targets in video.The perception process of dynamic video targets in the actual scene is very complicated.How the brain transits from searching to spotting,how to allocate resources and interact information in multiple brain regions are the basic problems to be solved in video target perception.Based on the EEG source imaging technology and time-varying brain network analysis method,this study reconstructs the cortical brain responses and maps time-varying brain network from searching to spotting,furtherly analyzes the dynamic regular pattern of information interaction.The results reveal that the process of task target perception roughly includes three phases: target searching,information integration,and decision-making where the brain activation,brain connection,and processing efficiency increased gradually;The control sources add the central cortex,inferior parietal gyrus,temporoparietal junction on the basis of superior frontal gyrus and middle occipital gyrus;The visual region and frontal parietal region are critical for the target decision-making;Target decisionmaking brain activity is concentrated below 10 Hz;There is a significant correlation between the searching network property and detection accuracy.2.Research on the perception mechanism of non-task targets in video.In the actual scene,non-task related things in the video have the characteristics of multi-occurrence,concurrency and randomnessa,which becomes the critical factor causing the high false alarm rate of video target detection.Exploring the perception differences between the non-task targets and the task targets is an important premise to solve this problem.This study introduces eye movement calibration technology to extract the EEG signals induced by non-task targets,reconstructs the source responses of non-task target perception,and analyzes its difference with task target perception.The results reveal that the brain activation and brain connection inspired by non-task targets were weaker than those of task targets;Information integration of non-task targets calls more resources in the left hemisphere,especially in the left superior frontal gyrus;Decision-making of task targets activated more the visual region and frontal parietal region,especially the left visual pathway.3.Research on an asynchronous detection method for video targets based on single-trial EEG signals.Different with the traditional static image targets,the dynamic targets in the video have no time anchor,which makes it difficult to accurately extract and analyze the single-trial EEG signal induced by the task target.This study constructs an asynchronous detection method for video targets based on single-trial EEG signals.In this method,a spatial filter is designed to reduce the dimension of latency jitter of multi-channel single-trial P3 signals to one-dimensional time series;An alignment algorithm based on minimum distance square error is proposed to eliminate the latency jitter of single-trial P3 signals in one-dimensional time series.The results show that the average accuracy of online single-trial P3 signal detection by the asynchronous detection method is 93%,and the F1 score is 0.74;Compared with the existing methods,the F1 score of this method is significantly improved by 0.03(p<0.01),which improves the timeliness of EEG-based video target detection.4.Research on an EEG individual transferring model for video target detection.The dynamics,uncertainty,and unpredictability of video targets aggravate the individual differences of EEG signals,resulting in the sharp decline of the detection performance and poor stability of the current EEG-based video target detection model in the cross-subject detection mode.This study proposes an unsupervised multi-source domain adaptation network(P3-MSDA)to solve the problem of individual transferring for EEG-based video target detection.The network presents a source domain individual selection method based on the P3 map clustering to select individuals with strong P3 maps as the source domain;Through feature extraction,domain confrontation,and category classification between source domain and target domain,the difference of individual features is reduced and the separability of feature categories is ensured;The target domain sample selector is designed to solve the problem of unbalanced sample classification.The results show that the average accuracy of P3-MSDA network in cross-subject detection is 84%,and the F1 score is 0.66;Compared with the existing methods,the F1 score of the network is significantly improved by 0.05(p<0.01),which improves the stability of EEG-based video target detection.5.Research on multi-brain collaborative brain computer interface technology for video target detection.The video target detection accuracy of single-brain BCI is still not high enough,and the c BCI has the advantage of performance enhancement.However,the existing c BCI technology lacks information interaction and learning guidance among multi-agents.This study proposes a novel c BCI system based on a mutual learning domain adaptation network(MLDANet).MLDANet takes P3-s SDA network as individual network unit,introduces mutual learning strategy,and establishes a dynamic interactive learning mechanism between individual networks and collaborative decision-making to highlight the advantages of multi-agent collaborative decisionmaking.The results show that the mutual learning strategy improves the group detection performance and individual network detection ability of MLDANet-c BCI system;In the crosssubject detection,the average accuracy of three-mind collaborative detection is 91%,and the F1 score is 0.73;Compared with the existing methods,the F1 score of the c BCI technology is significantly improved by 0.12(p<0.01),which improves the accuracy of EEG-based video target detection.