基于背景差分的视频车辆检测

发布时间：2018-06-18 20:30

本文选题：车辆检测 + 稀疏表示　；参考：《江苏科技大学》2016年硕士论文

【摘要】：基于视频的车辆检测是智能交通系统的基础和关键,也是计算机视觉领域的研究热点。随着研究的深入,近年来提出了许多新的车辆检测算法,逐步解决了检测过程中出现车辆阴影与背景扰动的问题,然而这些算法仍然存在着一些缺点和不足,如没有考虑雨雪雾霾等恶劣天气环境下视频图像的噪声和干扰问题,以及车辆检测算法的复杂度高所带来的实时性问题等。因此,本文针对当前基于视频的车辆检测算法中存在的噪声干扰和实时性问题,提出了具体的提高检测率和实时性的方法。主要的研究工作有以下两个方面:(1)针对视频中噪声影响车辆检测率的问题,深入研究了中值滤波、维纳滤波和小波滤波这三种常用的去噪算法,分析它们在实际应用中的不足,提出了一种基于稀疏表示的视频去噪方法。利用K-奇异值分解(K-Singular Value Decomposition,K-SVD)算法训练过完备字典,将噪声图像在过完备字典上稀疏分解,根据图像信号能够在过完备字典上稀疏分解而噪声不能稀疏分解的原理,去除噪声,恢复图像。与上述三种去噪方法相比,本文提出的去噪方法能够明显提高恶劣天气或环境下视频图像的信噪比,改善图像质量,有助于提高车辆的检测率。(2)围绕背景差分方法中如何建立高鲁棒性的背景模型问题,在深入分析和研究均值法、单高斯背景模型和混合高斯背景模型三种经典的建模算法的基础上,提出了一种基于混合高斯的多模态模型的优化方案。由于建模过程中所得到的所有模型都有可能是背景模型,模型权值较小的像素中有可能也包含了真实背景像素。因此,舍弃建模过程中模型匹配后的权值排序、累加值与阈值比较这两个计算步骤,从而达到既能建立更真实的背景模型又能降低算法的计算量的效果。最后,利用质心跟踪法完成多车辆的跟踪,并统计车流量。实验结果表明,本文提出的基于混合高斯的多模态建模的车辆检测算法具有抗干扰能力强、计算量低和检测效果好的优点。
[Abstract]:Vehicle detection based on video is the basis and key of Intelligent Transportation system (its), and it is also a hotspot in the field of computer vision. With the deepening of research, many new vehicle detection algorithms have been proposed in recent years, which gradually solve the problem of vehicle shadow and background disturbance in the detection process. However, these algorithms still have some shortcomings and shortcomings. Such as not considering the noise and interference of video image in severe weather such as rain, snow and haze, and the real-time problem caused by the high complexity of vehicle detection algorithm, etc. Therefore, aiming at the noise interference and real-time problems in the current video based vehicle detection algorithm, this paper proposes a specific method to improve the detection rate and real-time performance. The main research work has the following two aspects: 1) aiming at the problem that the noise in the video affects the vehicle detection rate, three commonly used de-noising algorithms, median filter, Wiener filter and wavelet filter, are studied in depth, and their shortcomings in practical application are analyzed. A method of video denoising based on sparse representation is proposed. Using K-Singular value decomposition (K-Singular value DecompositionK-SVD) algorithm to train over-complete dictionaries, the noise images are sparse decomposed in over-complete dictionaries. According to the principle that image signals can be sparse decomposed in over-complete dictionaries but noise cannot be sparse decomposed, the noise is removed. Restore the image. Compared with the above three denoising methods, the proposed denoising method can significantly improve the SNR and image quality of video images in severe weather or environment. It is helpful to improve the detection rate of vehicles. (2) focusing on how to establish a background model with high robustness in background difference method, this paper analyzes and studies the mean value method in depth. On the basis of three classical modeling algorithms of single Gao Si background model and mixed Gao Si background model, an optimization scheme of multimodal model based on hybrid Gao Si is proposed. Since all the models obtained in the modeling process are likely to be background models, it is possible that the pixels with small weights of the model may also contain real background pixels. Therefore, the weight ranking after model matching is abandoned, and the cumulative value is compared with the threshold in order to establish a more realistic background model and reduce the computational cost of the algorithm. Finally, the centroid tracking method is used to track multiple vehicles, and the traffic flow is counted. The experimental results show that the vehicle detection algorithm proposed in this paper based on hybrid Gao Si has the advantages of strong anti-jamming ability, low computational complexity and good detection effect.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：U495;TP391.41

【参考文献】