为保障核能的可持续发展和社会的安全稳定，核设施的安保问题越来越受到关注，中美国家元首连续参加四届核安保峰会，着重强调要重视核安保的发展。然而现有的核设施安保系统，均参照IAEA的标准体系，缺乏对于车辆的检查手段，难以防范爆炸物等违禁品的非法带入和核材料的非法带出。因此，清华大学核研院研发了“核设施出入口车辆综合检查系统”。该系统为门架移动式双投影检测系统，从两个不同方向对车辆进行透射投影，能够同时获取两幅投影图像。系统研究难点在于如何综合利用两幅投影图像获取可疑物体的位置、尺寸和密度等信息。本文基于双投影检测系统，对于透射截面为凸多边形的物体，提出综合两方向投影信息，分析透射所得灰度曲线特点，结合系统结构对其进行关联计算，得到目标物体的近似尺寸、位置、密度等特征，将此近似信息作为初值，使用梯度下降法得到上述特征的更精确结果，并通过物体的密度和衰减系数进行物性判别。首先，为验证该方法，在MATLAB平台上进行仿真，建立与实验系统一致的仿真检测系统，给定单一材质样块，获取灰度曲线。通过坐标化计算、迭代计算等步骤，得到目标物体位置、形状及吸收系数。对无噪声仿真数据的衰减系数计算结果表明，拐点反算方法和迭代方法的相对误差为1.68%和0.34%；对叠加噪声仿真数据，两种方法相对误差为2.11%和0.4%。迭代方法具有更高的准确性和鲁棒性。其次，利用双投影检测系统，对铝材质的立方体进行实际透射实验，选择其中某一截面进行形状及尺寸计算，并得到目标物体的吸收系数。为更接近实际检测情况，将饮水机水桶置于卡车货箱内进行实验测量。对卡车剪影后选取一系列水桶截面进行计算，重建出该待测物体的立体模型轮廓，并计算得到吸收系数进而得到目标物体密度，实验结果相对误差为7.8%。最后，仿真过程中基于实际透射检测装置的放射源及探测器布置结构建立仿真环境，得到一系列与实际情况相近的仿真数据。利用机器学习中的重要方向分类学习，挖掘仿真数据特征。通过仿真数据建立分类学习训练集，分别采用基于集成学习的袋装树方法和基于K近邻分类方法对部分数据进行训练，其余部分仿真数据作为测试集，实际工业数字辐射成像数据作为验证集对模型进行准确性验证。结果表明袋装树方法对基于仿真数据的辐射成像分类学习具有较好的效果。文末，根据现有工作对课题进行总结，并基于实验结果对未来可能开展的工作提出设想与展望。

In order to ensure the sustainable development of nuclear energy and stability of so-ciety, the security of nuclear facilities has been paid more and more attention. However, the existing security system of nuclear facilities refers to the IAEA standard system, which lacks vehicle inspection. It is impossible to prevent illegal import of prohibited substances or export of nuclear materials through these system. Therefore, in this paper a new non-intrusive method to inspect vehicles is proposed. Approximate size, position and den-sity of the objects can be obtained by analyzing the data obtained. Then, the objects are discriminated by the density and attenuation coefficient. Firstly, the simulation experiment is carried out. The position, shape and attenuation coefficient of target object can be obtained through coordinate calculation and iteration calculation. The results of attenuation coefficient calculation show that the relative error of inflection point inversion method higher. The iteration method has higher accuracy and robustness.Then, practical experiment of aluminum cube and water bucket are car-ried out based on the dual-projection inspection system. The shape and size of central slice are calculated, the attenuation coefficient and the density of target object are also obtained. The calculated results has high accuracy.Finally, due to the complex content and irradiation conditions of indus-trial radiation imaging images, it is consuming to obtain complete training set samples for classification learning by practical measurement. The method aims to obtain complete training set data accurately and quickly by using simula-tion data which is verified has a good effect on classification learning of ra-diography imaging.Due to the simplification of target object is differ from the practical inspection situation, using a more specific model to describe the target objects to be more close to the practical inspection situation and optimizing the al-gorithm in order to reduce the time of calculation and improve the accuracy of processing are the prospection work in the future research.