基于视觉感知的弱对比度车辆目标识别

发布时间：2018-03-13 21:20

本文选题：弱对比度车辆识别　切入点：选择注意　出处：《北京交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：由于造价低廉、易于应用,基于图像的车辆识别技术而成为近年来国内外研究的热点。国内外相关学者提出了很多富有建设性的方法并取得了一定成功,但是目前仍然存在环境适应性和鲁棒性差的缺点,且对于复杂交通场景和恶劣天气(如雾霾、雨雪等)下的弱对比度车辆目标难以获得令人满意的识别率,这已经严重制约了基于图像的车辆识别技术的发展。因此车辆识别,尤其是弱对比度车辆目标识别已经成为一项具有挑战意义和重要研究价值的工作。针对以上问题,本文借鉴人类的视觉感知原理建立了适应性强、鲁棒性好的车辆识别模型,并探讨建立联想机制模型用于弱对比度车辆目标的准确识别。论文的具体工作如下： 1、基于人类的视觉选择注意机制建立了双向驱动融合的注意模型用于车辆识别。该模型在基于bottom-up数据驱动的Saliency模型基础上,选择车辆目标最鲁棒的结构和形状特征建立两级知识库实现了top-down的任务驱动,在高层指导Saliency模型中的视觉选择注意过程,实现了任务驱动与数据驱动的融合。其中,在数据驱动过程中,利用谱分析方法和显著度函数代替了基于高斯金字塔的多尺度显著特征融合算法,提高了模型的实时性；在构建形状知识库时,利用格式塔知觉理论的相关原理建立了用于提取车辆目标闭合边界集的多目标分割模型。 2、探讨建立了用于弱对比度目标识别的联想机制模型。提出了具有优秀联想能力的绿色神经元交互联想网络,利用高维联想空间映射网完成了对于弱对比度目标不完整特征的模式异联想,并通过解联想映射网实现了目标的自联想功能；同时建立神经调节函数和神经交互函数模拟了生物神经信号传导过程中神经元的交互作用,使联想网络具有更快的收敛速度。在此基础上,本文借鉴人类视皮层中的WHAT通路将大脑皮层的联想功能合理抽象为联想产生、联想匹配和综合分析的层次化模型,从而构建了能有效识别弱对比度目标的联想机制模型。通过仿真实验得出,双向驱动融合的注意模型对于只有清晰车辆目标的样本集的识别率为90.4%,误识别率为4.9%；对于包含弱对比度车辆目标的测试样本集的综合识别率为76.4%,综合误识别率为4.8%；引入联想机制模型后系统的综合识别率为88.5%,综合误识别率为4.9%。实验结果表明,对于清晰车辆目标,双向驱动融合的注意模型具有很高的识别率,且鲁棒性强；引入联想机制模型能在保证系统鲁棒性的基础上显著提高系统对于弱对比度目标的识别能力。
[Abstract]:Because of its low cost and easy application, the image-based vehicle recognition technology has become a hot topic in recent years. Many constructive methods have been put forward by domestic and foreign scholars and some success has been achieved. However, there are still shortcomings of poor environmental adaptability and robustness, and it is difficult to obtain satisfactory recognition rate for vehicle targets with weak contrast in complex traffic scenarios and severe weather (such as haze, rain and snow). This has seriously restricted the development of image-based vehicle recognition technology, so vehicle recognition, especially the weak contrast vehicle target recognition, has become a challenge and important research work. In view of the above problems, this paper builds a vehicle recognition model with strong adaptability and good robustness based on the principle of human visual perception. This paper also discusses the establishment of associative mechanism model for the accurate identification of vehicle targets with weak contrast. The specific work of this paper is as follows:. 1. Based on the human visual selective attention mechanism, a bidirectional driving fusion attention model is established for vehicle recognition. The model is based on the bottom-up data-driven Saliency model. Selecting the most robust structure and shape features of vehicle targets, a two-level knowledge base is established to realize the task driven of top-down, and the visual selection attention process in the high-level Saliency model is guided by the fusion of task driving and data driving. In the data-driven process, spectral analysis method and saliency function are used to replace the multi-scale salient feature fusion algorithm based on Gao Si pyramid, which improves the real-time performance of the model. Based on the related principle of Gestalt perception theory, a multi-objective segmentation model for extracting closed boundary sets of vehicle targets is established. 2. The association mechanism model for weak contrast target recognition is established, and a green neural interactive association network with excellent association ability is proposed. Using the high-dimensional associative space mapping net, the pattern heterodyne association for the incomplete feature of the weak contrast target is completed, and the self-associative function of the target is realized through the de-associative mapping net. At the same time, the neural regulation function and the neural interaction function are established to simulate the interaction of neurons in the process of biological nerve signal transduction, which makes the associative network converge faster. Based on the WHAT pathway in human visual cortex, the association function of cerebral cortex is reasonably abstracted into a hierarchical model of association generation, association matching and comprehensive analysis, and a model of association mechanism which can effectively identify the target of weak contrast is constructed in this paper. The simulation results show that, The attention model of bidirectional driving fusion has a recognition rate of 90.4 for a sample set with only clear vehicle targets and a false recognition rate of 4.9.The synthetic recognition rate for a test sample set containing a weak contrast vehicle target is 76.4, and the comprehensive error recognition rate is 76.4. After introducing the associative mechanism model, the comprehensive recognition rate of the system is 88. 5 and the comprehensive error recognition rate is 4. 9. The experimental results show that, For clear vehicle targets, the attention model of bidirectional driving fusion has a high recognition rate and strong robustness, and the associative mechanism model can significantly improve the recognition ability of weak contrast targets on the basis of ensuring the robustness of the system.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP391.41;U495

【参考文献】