Study On Abnormal Acoustic Event Detection In Urban Security

In recent years,with the development of the China’s economy and the rise of emerging cities,more and more threats and challenges are being faced by national security,and its’ situation is very grim.From the domestic perspective,the conflicts,such as the expansion of the urban population,the widening of the gap between the rich and the poor,and the differentiation of social interests,have entered into an unprecedented active period.What’s more,from the international level,social public order is frequently disrupted by terrorist attacks.In order to solve the problems of the urban security surveillance system,such as poor timeliness,low detection rate and large limitations,the abnormal acoustic event detection(AAED)system is proposed by this thesis.And the AAED system is adopted to supplement the urban security surveillance system,and research is carried out from three aspects of signal pre-processing,feature extraction and classification recognition.Then these three aspects have been proposed for improvement respectively.Pre-processing technology of continuous signals received by sensors is an important basis for acoustic event detection.First the abnormal sound signals are processed by the traditional pre-processing methods,which includes framing,windowing,and pre-emphasis,etc.At the same time,the double-threshold endpoint detection technique based on short-term energy and zero crossing rate is proposed to cut the signal.The simulation results prove that the technology can not only effectively reduce the computational overhead of the AAED system,but also improve the real-time performance of the system,and provide a strong guarantee for subsequent classification and identification work.Extracting time-frequency features of abnormal sound signals based on orthogonal matching pursuit algorithm(OMP)is proposed in this thesis.The approach constructs a Gabor dictionary based on the critical frequency band of human ears,uses the OMP algorithm to sparsely decompose the signal,selects the optimal several important atoms from the Gabor dictionary,extracts their scale,frequency and displacement parameters to obtain OMP time-frequency feature.Simulation results show that OMP time-frequency features have good performance for abnormal acoustic event detection.Furthermore,the problems of low recognition accuracy and poor robustness for the surveillance system can be addressed by combining them with traditional acoustic features in practical applications.Since the random forest is the best single classification algorithm for AAED,combining it with ensemble learning,an integrated random forest model based on weighted soft voting is proposed in this thesis.The model integrates multiple random forests,uses weighted soft voting as the ensemble strategy,and optimizes the parameters of the proposed model based on intelligent optimization algorithm.At the same time,for the imbalance data faced by the ensemble model,two solutions based on the Bagging algorithm and the Bagging ensemble variation(BEV)algorithm is proposed in this thesis.The simulation results show that compared with the traditional single classification algorithm,the proposed model achieves the purpose of improving the recognition rate of abnormal acoustic events.As the strong global search ability and fast convergence speed of the particle swarm optimization algorithm(PSO),the model with higher classification accuracy based on parameters optimization of the PSO have been obtained.In addition,the data equalization scheme based on the BEV can effectively improve the detection rate of the minority class data,and enhance the generalization ability of the model while balancing the abnormal sound data.