School Violence Detecting Algorithm Based On Multisensor Data Fusion

With the development of the information times and network media technology,people have increasingly rich channels for receiving various kinds of news.The school violence incidents that took place in the quiet corners of the campus have gradually attracts people's attention.With the popularity of the Internet and the various kinds of self-media,the riotous and complicated online world,as well as all kinds of violence and vulgar content have seriously affected young teenagers with immature minds.Many students are full of strong curiosity about the vulgar content they are exposed to.Imitate the mentality,thus making the campus violence emerge in an endless stream,which seriously affects the student's campus life and social atmosphere.However,because violent incidents are often secretive and most students cannot report to teachers or parents in a timely manner due to psychological fear,the result is a vicious cycle of school violence,leaving an indelible psychological shadow on students' minds and seriously affecting the physical and psychological development of victims.With the development and popularization of all kinds of wearable smart devices,the use of wearable devices built-in sensors has also been developed for action recognition.This thesis proposes an active detection of campus violence based on multi-site sensor data fusion.Compared with single-site sensor identification,multi-sensors can collect more comprehensive body movement data,and achieve redundant and complementary effects through information fusion to get more accurate identification of complex violence.The active violence detection system can prevent students from being bullied and not be discovered.It can timely and effectively warn students about campus violence incidents,which is of great significance to promote the harmonious development of campus and students' physical and mental health.Based on the complexity of the violence on campus,this thesis synchronizes and pre-processes the data of multiple location sensors,the features are extracted based on the analysis of the violent action and the daily action.The defects of the Relief-F algorithm are improved.Then a decision tree-RBF neural network multi-level classifier is constructed to identify the actions in the foundation of the improved Relief-F feature selection algorithm.and the Posterior-adapted Class-based fusion algorithm is adopted at the decision-making level to the waist data and legs.Finally,the recognition rate of the violent action is 84.4%,the normal action recognition rate reaches 97.0%,compared with single-site data,the recall rate was improved by 5.0%.Then,the LDA dimensionality reduction algorithm is introduced to the problem of high complexity of the algorithm,which reduces the feature dimension to 8 dimension through the LDA reduction algorithm and reduces the system operating time by about 51% while ensuring the overall system performance,and the recognition rate of the specific complex action is raised by about 10%.Through the LDA dimension reduction laid the foundation for future hardware implementation.Finally,for the problem that the recognition results of different parts of the decisionmaking layer fusion process cannot be correctly merged,based on classical D-S theory,a new probability distribution function is designed to modify the original evidence model and construct a new fusion rule to improve the D-S algorithm.The algorithm can solve the fusion conflict problem primely.Finally,the recognition rate for non-violent actions reaches 95.0%,and the recall rate for violent actions increases to 90.0%,thus reducing the system's missing alarm rate.