视频侦查中多摄像头下嫌疑目标同一的概率研究  

Probabilistic Approach to Identifying Same Suspected Target from Multiple Cameras in Video Investigation

作　　者：黎智辉[1] 谢兰迟 吕游[2] 王桂强[1] LI Zhihui;XIE Lanchi;LÜYou;WANG Guiqiang(Institute of Forensic Science,Ministry of Public Security&National Engineering Laboratory for Forensic Science,Beijing 100038,China;Division of Crime Investigation,Tianjin Municipal Public Security Bureau,Tianjin 300384,China)

机构地区：[1]公安部物证鉴定中心,现场物证溯源技术国家工程实验室,北京100038 [2]天津市公安局刑侦局,天津300384

出　　处：《刑事技术》2022年第1期24-34,共11页Forensic Science and Technology

基　　金：国家重点研发计划(2017YFC0803506);公安部技术研究计划(2019JSYJA06)。

摘　　要：目的视频侦查中的时空轨迹信息是很多案件侦破的关键,但在证据准备和法庭质证阶段,如何使用这些信息目前并没有方法依据。本文针对视频侦查中轨迹信息的证据转化问题,开展理论分析,建立概率方法。方法结合视频侦查的典型过程,定义了摄像头下人物的人体、衣着和交通工具等特征;对不同的特征根据其出现的可能性进行针对性的概率估计以得到概率上限,对多摄像头下嫌疑人作案和逃窜过程的概率进行表示;基于图模型构造贝叶斯网络对轨迹追踪问题进行建模,并对模型利用贝叶斯网络特性进行求解,以得到似然比结果。结果结合视频侦查场景给出了问题中似然比的公式和近似计算方法。1)通过分解提出时段、时序和方向假设建立了一种时空特征概率模型分析方法;2)提出了视频图像中衣服、人体、交通工具特征的概率分析思路框架;3)得到了假设条件下的似然比结果;4)给出了假设条件对概率计算的影响,讨论了实际应用中后验比估计的方法及方法的使用限制因素。结论本文建立的理论分析方法首次为视频追踪结果的证据应用提供量化参考模型,从概率的角度解释了视频轨迹追踪结果为什么可以成为证据,有助于提升视频追踪证据的证明力:基于视频中能观察到的常见的十二个特征(体态、上衣颜色、上衣纹理、上衣款式、裤子款式、裤子颜色、鞋子款式、鞋子颜色、头发长度、摩托车款式、摩托车颜色、摩托车行李箱样式)的条件,对于具有相同特征的目标依次出现在三个摄像头下(摄像头之间没有分叉路口)的案例,计算得到的似然比(第一个摄像头下的目标与第三个摄像头下的目标为同一目标和不同目标条件下,出现给定特征的可能性之比)下限超过10;量级。Objective The spatial-temporal track information is key to find the suspected target in video investigation for many cases,yet lacking a methodic basis on how to use such information in the stage of evidence preparation and court testimony.Thus,a probabilistic approach was here to try for the problem to solve with evidence transformation of the target’s track information in video investigation.Methods The characteristic features were specified for human body,clothes and transportation vehicle(motorcycle in this essay)to go past the surveillance cameras under typical process of video investigation so that the different features were estimated of their presence probability plus its upper limit.Accordingly,the probabilities were to express the course of suspect’s committing crime and escaping under multiple cameras,therewith having converted the probability-based track query from a graph-representing model to one Bayesian network whose characteristics were hence able to utilize to have the likelihood ratio estimated.Results The formula and approximate calculation method were given about the likelihood ratio for the concerned problem put under video investigation scene,along with 1)a feature-based spatial-temporal probability model plus analysis being established through decomposition of the proposed time segment/range and direction hypothesis;2)one probabilistic analysis plus thought framework being proposed about the characteristics of clothes,human body and vehicle in video image;3)the probability ratio(i.e.,likelihood ratio)results being obtained with hypothetical conditions;and 4)the influence being given of hypothesis on probability calculation,meanwhile both the method of posterior ratio estimation and its limiting factors being discussed for practical application.With the condition of twelve commonly-seen features(body shape,coat color/texture/style,pants style/color,shoe style/color,hair length,motorcycle style/color/luggage style)observed under three consecutive cameras(fixed onto a simply non-divergent

关 键 词：视频侦查 概率模型 贝叶斯网络 轨迹追踪 

分 类 号：D918.2[政治法律—法学]