【作者】 胡满江；

【导师】 葛如海；

【作者基本信息】 江苏大学 ， 车辆工程， 2014， 博士

【摘要】 为提高安全带佩戴率,论文从三点式安全带使用过程中存在的问题出发,分析三点式安全带的不同佩戴方式对乘员损伤的影响,利用车载机器视觉对其进行识别,提出安全带识别评价方法,构建满足实时性和高精度要求的识别模型,实现了嵌入式车载安全带监控系统设计。论文研究内容包括：1)三点式安全带使用情况调查与分析。针对我国典型城市安全带的使用状况展开调查,结果显示,安全带使用过程中存在两类问题：不规范佩戴安全带和卷曲佩戴安全带。其中,导致安全带佩戴提示系统失效的不规范佩戴行为包括：单独使用安全带带扣、预佩戴和仅系肩带；卷曲佩戴包括：肩带卷曲、腰带卷曲和严重卷曲。2)三点式安全带的不同佩戴方式对乘员损伤影响。采用MADYMO构建乘员约束系统模型,并验证模型的有效性。运用该模型进行三点式安全带在不同佩戴方式下的乘员损伤分析,结果表明：车辆碰撞时,预佩戴安全带的乘员将从座椅位置飞出；仅系肩带将产生明显的下潜、滑移运动。上述情况均导致乘员受到致命性损伤。相比正确使用三点式安全带,卷曲佩戴将导致乘员各项损伤指标明显上升。因此,对三点式安全带不同的佩戴方式进行识别,为充分发挥安全带应有的保护作用,具有积极意义。3)车载视频监控平台搭建及安全带图像采集试验设计。搭建安全带图像采集试验平台,在CCD传感器性能参数、红外补光装置和特殊材料安全带等方面展开具体研究；设计并实施不同光线环境下的车辆行驶试验,采集获得安全带不同佩戴方式下的乘员图像信息；研究适用于安全带识别的图像预处理技术,为构建安全带识别的多特征参数模型奠定基础。4)构建满足实时性要求的安全带识别模型。针对安全带在线检测的要求,采用主成分分析法降维后的安全带空间参数特征作为输入向量,选用BP神经网络作为分类器,构建基于BP神经网络的安全带识别模型,以满足实时性要求。引入遗传算法(GA)对其内部参数进行优化,构建基于GA-BP神经网络的安全带识别实时性模型,以满足准确性要求。通过硬件在环测试(HIL)和模型在环测试(MIL)验证了模型的实时性和准确性。5)构建满足高精度要求的卷曲佩戴识别模型。提取安全带结构参数的统计特征值作为输入向量,选用支持向量机(SVM)作为模型核心分类器,以交叉验证方法对内部参数进行选择,采用粒子群算法(PSO)对其进行优化,构建基于PSO-SVM的高精度识别模型,并进行软件在环测试以验证代码的有效性,从而将其应用于高精度离线检测之中。6)实现了嵌入式车载安全带监控系统的设计。从分析车载系统对软硬件的性能要求着手,探讨了DSP内核高速处理数据的特点以及ARM内核控制和管理的功能,选择ICETEK-DM642-B评估板作为硬件平台,实现了基于嵌入式技术的安全带识别系统的设计；完成系统功能的总体设计和功能模块的软件设计,并进行程序优化。论文研究的创新点如下：1)提出基于车载机器视觉识别安全带的方法,搭建嵌入式车载安全带识别系统平台；2)揭示了三点式安全带的不同佩戴方式(不规范佩戴和卷曲佩戴)与乘员损伤之间的关系；3)提出安全带识别的实时性和准确性评价方法；4)构建基于GA-BP神经网络的安全带识别实时性模型；5)构建基于PSO-SVM的卷曲佩戴识别高精度模型。更多还原

【Abstract】 In order to increase the wearing rate of seat belt, influence of different3point seat belt wearing patterns on occupant injury mechanism is analyzed, due to the existing problems in seat belt utilization. An evaluation criteria of seat belt identification is proposed by identifying seat belt with vehicle-mounted machine vision. An identification model is built to meet the requirements of real time and high precision, embedded vehicle-mounted seat belt identification system is designed and realized. The main research content is as follows:1) Investigation and analysis of3point seat belt utilization. Investigations were launched in typical cities in China. The results show that, problems existing in using seat belt mainly are:non-standard wearing and curly wearing. Non-standard wearing patterns that cause the failure of seat belt identification system includes:using only seat belt buckle, pre-wearing seat belt and using only shoulder belt. Curly wearing patterns includes:shoulder belt curling, waist belt curling and severely curling.2) Influence of different3point seat belt wearing patterns on occupant injury mechanism:the occupant restraint system model is built based on MADYMO software, and its validity is verified. Simulation on different seat belt wearing patterns is conducted. The simulation results demonstrate that:in the event of vehicle crash, the occupant without seat belt will be thrown out of the seat; the occupant wearing only the shoulder belt will have obvious diving and gliding movement. Compared with the situation when three-point seat belt are used correctly, curly wearing will cause a significant increase in occupant injury criterion. Therefore, identifying different wearing patterns has a positive impact on the protection function of the seat belt.3) Vehicle-mounted video monitoring system platform construction and test design. Vehicle-mounted video monitoring system platform is built, research is conducted on CCD sensor performance parameter, infrared light device and seat belt of special material. Vehicle field driving tests in different lighting environments are designed and conducted, occupant image information of different seat belt wearing patterns is collected. Image preprocessing technology appropriate for seat belt identification is studied in order to lay the foundation of building multiple characteristic parameter model for seat belt identification.4) Real-time seat belt identification modeling. Aimed at the requirement of seat belt on-line test, space parameter characteristic after dimensionality reduction based on PCA (principal component analysis) is selected as input vector, BP neural network is chosen to be the classifier to meet the requirement of real-time identification, genetic algorithm (GA) is brought in to optimize the internal parameters to improve accuracy. At last, a real-time seat belt identification model based on GA-BP neural network is built. The real-time performance and accuracy of this model are verified through hardware-in-loop (HIL) and model-in-loop (MIL).5) Curly wearing identification model with high accuracy is established. Statistical characteristic value of seat belt structural parameters is extracted as input vector, support vector machine (SVM) is chosen as the core model classifier, internal parameters are selected by cross validation method and optimized by particle swarm optimization (PSO), an identification model with high precision based on PSO-SVM method is built. The validation of the code in this model is verified through software-in-loop test, so that it can be applied to high accuracy offline testing.6) Embedded vehicle-mounted seat belt identification system is designed and realized. The requirement of vehicle-mounted system on hardware and software is analyzed, high-speed data processing in DSP kernel, control and management fuction of ARM kernel are studied. ICETEK-DM642-B evaluation board is chosen as the hardware platform, and the seat belt identification system is realized based on embedded technology. System function overall design and function module software design are achieved, system program is optimized.The innovations of this paper are as follows:1) A method of seat belt identification based on machine vision is proposed for the fist time, embedded vehicle-mounted seat belt identification system platform is built.2) The relationship between seat belt wearing patterns (non-standard wearing and curly wearing) and occupant injury is revealed. 3) The criterion to evaluate real-time performance and accuracy in seat belt identification is proposed.4) The real-time seat belt identification model based on GA-BP neural network is built.5) The curly wearing identification model with high accuracy based on PSO-SVM is built.更多还原