Research On Moving Ground-Target Perception And Recognition Method Based On Vibration

The perimeter security of important areas is a key issue that relates to the national territorial security and the security of enterprise and personal property.China has a vast territory and multiple borders countries.The illegal behaviours(e.g.,smuggling drugs)in border areas are banned but still repeats in some areas.In recent years,the security and protection in the border of our country have faced more severe challenges under the influence of epidemic,which leads to a serious threat to the security and the health of our people.Meanwhile,the security of important areas(e.g.,military bases,prisons,ancient tombs and base stations)is also with great importance.The illegal behavior caused by illegal targets in the afore-mentioned areas will lead to a severe threat to the stability of society and the life of people.Existing perimeter security systems are mainly based on camera,which is not easy to be concealed and difficult to conceal.Moreover,camera is difficult to work in the wild and in environments that have poor infrastructure.Therefore,it is urgent to build a security system by adopting a target perception method that can be deployed in harsh environment.Such security system can realize the perception of illegal targets.The system can detect and recognize illegal targets,which will provide a scientific management and intelligent protection for borders,and avoid the occurrence of intrusion and destruction.When targets(i.e.,pedestrians and vehicles)move on the ground,they will generate vibration signals that propagate through the ground.Such vibration signal is an effective means for detecting and recognizing moving ground-targets in perimeter security systems.Meanwhile,the vibration sensors are small enough to be unnoticeable from people and need little power consumption,which can run a long period of time in the field area.Moreover,the workinng characteristics of vibration sensors are hardly influenced by weather and complex terrain.Therefore,vibration sensors are suitable for collecting the vibration signals generated by moving ground-targets in various environments.However,existing moving ground-target perception and recognition methods based on vibration signals has high false alarms and false alarms,low recognition accuracy and poor environmental adaptability.These problems badly influence the performance of perimeter security systems.Aiming at existing problems in the research of moving ground-target perception and recognition,this paper studies the generation,propagation and attenuation of vibration signal,and proposes signal preprosseing method,moving ground-target detection and recognition method.The researches and innovations of this paper are summarized as follows.(1)This paper proposes a vibration signal preprocessing algorithm based on variational mode decomposition,which can reduce the influence of environment noise on the perception and recognition of moving ground-targets.The original vibration signal is processed by variational mode decomposition and the is decomposed into multiple modes which have different center frequencies and separated characteristic time scales.Then,an adaptive mode division strategy defined by functions is proposed to divide the decomposed modes.Specifically,Euclidean distance and correlation coefficient are combined to construct the adaptive mode division strategy.According to the spectral distribution of the decomposed modes,the modes are divided into three types: effective component dominated modes,noise component dominated modes and noise-only modes.Such division of the mode achieves a more complete mode selection,which realizes a more systematic selection of signal components.Finally,the three modes are processed separately and the denoised signal is reconstructed.The experimental results show that the proposed preprocessing method can effectively improve the signal-to-noise ratio,and is suitable for vibration signals generated by pedestrians and vehicles.Specifically,the proposed algorithm can effectively suppress the noise components in pedestrian vibration signals,and the peaks of the processed waveform are smooth and retain the target features in the original signal.For vehicle vibration signals,the proposed algorithm can suppress the glitch signal caused by noise,which indicates a good signal smoothing effect and noise reduction ability.(2)This paper proposes an adaptive moving ground-target detection algorithm based on multiple features,which can solve the problem of high false positive rate and high false negative rate of existing methods.First,studing the different representation between the target’s signal and the environment noise,and analyzing the best combination of the effective detection features of the signal are selected by combining random forest and out of bag score.Such combination of multiple features can avoid the false alarm caused by single feature and reduce the influence of impulse signal on the performance of signal detection method.Moreover,it can also improve the reliability of the detection method.Then,the detection problem is transformed to a classification problem.The target detection model is trained by the support vector machine to distinguish the target vibration signal and the environmental noise.Finally,this method uses short-time average and long-time average to construct a feature transformation strategy to optimize the adaptability of the detection method.Specifically,calcutes the time-domain energy ratio between long-time window and short-time window to judge whether the real-time input feature vector is needed to be transformed,and maps the features of current signals and training signals in training dataset to a same feature space using linear transformation.The experimental results show that the proposed algorithm has good detection performance,the accuracy of the proposed method detects pedestrian targets is 97%,improving the accuracy of existing detection methods by 15%.The accuracy of the detecting vehicles is 94%,improving the accuracy of existing detection methods by 11%.Moreover,the detection accuracy of the proposed algorithm in the new environment is 98.9%,improving the accuracy of existing detection methods by 26.8%,which indicates a good environmental adaptability.(3)This paper proposes a ground moving target recognition algorithm based on neural network and domain adaptation,which can solve the problem that the existing recognition methods has poor applicability in new environment.Aiming at the distinguishing characteristics of pedestrian and vehicle vibration signals,a neural network is used to extract features from the signal time-domain waveform and timefrequency graph to build a target recognition model.Specifically,the low-frequency part of the short-time Fourier transform(STFT)of signals is used to construct a twodimensional feature matrix.CNN that is composed of small convolution kernel and fewer convolution layers is used to construct feature extraction network.The proposed model has high recognition accuracy and takes less training and computing time,which can recude the power consumption of the system.Furthermore,the semi-supervised learning of the recognition model is constructed by using the domain adaptive algorithm,which transfoms the recognition model from source domain to target domain.A joint loss function composed of cross entropy,focal loss and Maximum mean discrepancy distance is constructed to constrain model training.Such combination improves the ability of feature extractor to extract domain-invariant features of the source domain and target domain,and improves the accuracy of recognition model in new environment.The experimental results show that the accuracy rate of the basic model proposed in this paper for identifying pedestrians,wheeled vehicles and tracked vehicles is 96.4%,improving the accuracy of existing recognition methods by 1.1%.The judgment time of single calculation of the model is 0.8ms,improving the calculation time of existing methods by 3.2ms.The average recognition accuracy of the algorithm transforms across multiple datasets is 93.4%,improving the accuracy of the model without domain adaptation by 29.4%,which indicates a good transferability and a good recognition performance.