Research On Human Action Recognition Exploiting Through-wall Radar

Through-wall human action recognition is of great value on the application in many fields,such as urban warfare,anti-terrorism combat,stability safeguard,and law enforcement.By exploiting the capability of penetrating non-metal obstacles for the electromagnetic wave with low frequencies,through-wall radar is afforded the ability to collect the action echoes originated from the hidden human target in a real-time manner.Based on these echoes,a series of features corresponding to different human action types are available to be extracted,and their dissimilarities are capable to be analyzed,enabling the recognition of the human action types in real time.However,because of the intricatenesses of the electromagnetic circumstances as well as the low working frequencies of the through-wall radar system,the hidden human action features can not be extracted in a high-resolution way.Moreover,with the factors of the temporal and spatial multifariousness of the hidden target actions,the recognition performance of the hidden human actions significantly degrade on the aspects of accuracy,timeliness,and robustness,rendering the through-wall human action recognition problem challenging.To surmount the aforesaid issue,this dissertation investigates the description and the characterization of the through-wall human actions,develops the real-time recognizing approach for time-varying hidden target actions with variable temporal lengths,and explores the robust recognition method under the multi-view condition.The major innovations are as follows:1.A method for describing the hidden human actions based on short-time feature state transition is proposed.By decomposing the action into a sequence consisting of a number of short temporal slices containing instantaneous information of human postures,the posture features are capable to be extracted at each instance,and are able to be associated across multiple temporal slices.The proposed method conquers the defects of the global action description approach on the non-real time aspect as well as the fixed temporal scale character,enabling the reasonable description of the through-wall human actions with various temporal length.2.A method for characterizing through-wall human actions based on range-profile embedding technique is proposed.By lowering the dimensionality and clustering the features of the range profile,the short-time action features of hidden human actions are able to be extracted effectively,implementing the temporal action semantic association among the short-time posture feature slices.3.A method for recognizing hidden human action types based on recurrent neural network is proposed.By constructing a gated recurrent unit(GRU)network that exploits the forgotten and memorable characters over the dynamic action information at each temporal instance,the through-wall human action recognition is able to be implemented in a real-time manner theoretically.4.A method for evaluating the hidden human action recognition model based on normalized harmonic weighted intersection over union(NHW-IOU)is proposed.This method overcomes the defects of the accuracy principle on stability and real-time character of the action recognition model,enabling the reasonable evaluation of the through-wall human action recognition model.5.A method for multi-view through-wall human action recognition based on ensemble learning is proposed.By effectively fusing the action features acquired from the radar located at different views,the hidden action recognition is implemented in a more robust way,addressing the unstable recognition problem under the single view situation.All the aforementioned work has been validated by real data.