Key Technology for Big Visual Data Analysis in Security Space and Its Applications

Abstract

Seeing (surveillance) but not understanding (poor security performance) common problem in most security systems. Although in recent years significant progress has been achieved in biometric identification technology, the progress of single technologies does not dramatically improve the overall performance or solve the system-level problems of social security. Currently in addition to improvements of single technology, the following system-level technical bottlenecks must be solved to improve the overall performance of social security: 1. Blind spots of perception exist in surveillance area of due to the non-overlap of surveillance sensors, 2. The performance of single identification technology decreases sharply in complex surveillance scenarios such as poor lighting conditions or disguise, 3. Traditional warning technologies become invalid due to the multi-stage non-stationary evolution feature of complex events. Three challenges listed above closely relate to three scientific problems in the analysis technology of big visual data on three levels: sensing data, identification technology and pattern recognition. Our study aims at 1) exploiting the complete mapping mechanism between physical space and multivariate sensing spaces to fill-in the blind spots of sensing data, 2) exploring the correlative mechanism of multi-modal objects in multivariate sensing spaces to improve the analytical performance from single identification technology, 3) studying the spatial-temporal evolution mechanism in the entire lifecycle of complex events to extend pattern recognition from local space and time. The security space the physical space with comprehensive protective ability, which includes ubiquitous sensing, reliable identification, penetrating trend analysis and approaching danger warning at any time, at any location, for any object and for any behavior. Aiming to build the theory of security space, this study divided into three levels: information acquisition and perceptional computation, scene analysis and evolution prediction, resource scheduling and system applications. Then it further divided into five tasks: 1) task 1: visual object recognition and big data based identification, 2) task 2: situational awareness of groups and multi-scale revolution, 3) task 3: semantic analysis of scenes and correlative computation in multivariate spaces, 4) task 4: big scale visual retrieval and security risk analysis, 5) task 5: Warning system of the security space and its applications. Task 1 corresponds to the first level because sensing and identification of objects the basis for analysis. Task 2 and 3 correspond to the second level since group behavior, events, scene and their evolution are crucial for the prediction and warning of security events. Task 4 and 5 correspond to the third level of this study in order to develop the high performance computing platform, to conduct system evaluation and application demonstration. Big visual data contains massive high-dimensional sensing data, implying the complicated relationship among social objects. In fact, in the world of data, the spatial-temporal relationship between the big data objects more essential than the causal relationship, and these private and implicit relationships compose the core values of the big data social analysis. Only the analysis of individuals, groups and scenes in big visual data are based on the core element of social security analysis, that is social structure and social activities, can it supports the strategic transference of urban security system from investigation afterwards to warning in advance. The overall purpose of this study to build the big data analysis system in security space, realizing the intelligent big visual data system which supports data analysis of hundred billions of feature data, billions of image data, millions of visual sensing terminals. The expected achievements on the warning and protection system for large spaces can reach the international leading level. Based on the achievements above, the project plans to develop 10 intelligent big data analysis products of 3 categories, and the expected benefits of industrialization promotion can reach 100 billion, which promotes the upstream and downstream industry to realize economic benefits 3 billion. Also we strive to become the internationally leading industry in the field of big security data analysis.