智能交通系统中视频目标检测与识别的关键算法研究

发布时间：2018-05-18 18:47

本文选题：智能交通 + 目标检测　；参考：《华南理工大学》2014年博士论文

【摘要】：视频目标的检测、识别是目前智能交通和计算机视觉领域中的一个重要研究方向。但是，由于检测和识别环境下存在背景复杂、光照变化、目标遮挡等原因，导致该应用仍面临着许多困难，检测和识别的鲁棒性及准确性都有待进一步提高。本论文对视频目标检测和识别中的几个关键问题进行了研究，主要包括：复杂场景下目标与背景、阴影的准确分割；对提取的前景目标准确分类；复杂背景下的目标识别。针对这些问题，本论文提出了相应的解决方法。具体工作如下： 1.提出了一种基于自适应模糊估计的背景建模方法。该方法从函数估计的角度对背景进行建模，并采用TSK模糊系统作为估计函数。为了训练函数估计算子，分别使用粒子群优化（PSO）算法和递归最小二乘估计（RLSE）算法来优化模糊系统的前件参数和后件参数。为了有效估计背景，将前景像素看作背景像素的异常样例，并提出了异常样例的去除方法，然后用去除后的结果去训练模糊估计算子。该方法在动态背景、光照变化、摄像机振动等环境下都具有较高的运行效率和检测效果。 2.提出了一种基于模糊积分的运动阴影检测方法。在提取前景区域的基础上，选择颜色和纹理作为阴影检测的特征，并分别定义了这两种特征的相似性和重要性测度函数，然后通过Choquet模糊积分将这两种特征融合，实现阴影和前景目标的分类，最后通过后续处理，找到真正的阴影区域。 3.提出了一种基于JointBoost I2C距离度量的目标分类方法。针对经典I2C距离计算量大且易受噪声干扰等不足，首先提出了一种原型特征集的生成方法，该集合中的样本数量较少，但更具有代表性，计算测试图像到该原型特征集的距离花费较少时间；然后借助JointBoost算法的思想，联合多个I2C距离度量生成一个强分类器；最后还提出了一种将空间信息融合到强分类器的方法。实验证明，该方法在前景目标和图像分类实验中，具有更高的分类性能。 4.提出了基于特征码本树和能量最小化的目标识别方法。该方法考虑了特征的空间位置信息和特征之间的空间关系，集成了目标检测和目标识别。首先从目标图像提取的大量特征中过滤掉噪声特征；然后对单特征和空间上邻近的串联双特征分别使用层次k均值聚类算法构建特征码本树，，利用树模型可以实现特征快速定位和分类；最后建立一个能量函数来融合单、双特征码本树的类别概率匹配结果，并通过在测试图像中寻找滑动窗口所在区域的能量最小化来确定所属类别目标的位置。 5.提出了基于优化Hough森林代价损失的目标识别方法。首先在充分利用训练图像中对象位置是已知的基础上,提出了改进的偏移量不确定性度量方法；其次借助Boosting算法的思想，学习图片块样本和目标对象样本的自适应权重分布，并分别优化用于构造随机树和Hough森林的代价损失函数；最后根据图片块样本的权重分布，提出了改进的类标志不确定性度量方法。基于Hough森林的代价损失函数，还提出了随机树权重的学习方法。
[Abstract]:The detection and recognition of video targets is an important research direction in the field of intelligent traffic and computer vision. However, because of the complicated background, illumination change and target occlusion in the detection and recognition environment, the application still faces many difficulties. The robustness and accuracy of detection and recognition need to be further improved.
In this paper, several key problems in video target detection and recognition are studied, including: the target and background of the complex scene, the accurate segmentation of the shadow, the accurate classification of the foreground object and the target recognition under the complex background.
1. a background modeling method based on adaptive fuzzy estimation is proposed. This method models the background from the angle of function estimation and uses the TSK fuzzy system as the estimation function. In order to train the function estimation operator, the particle swarm optimization (PSO) algorithm and the recursive least double multiplicative estimation (RLSE) algorithm are used to optimize the pre fuzzy system. In order to effectively estimate the background, the foreground pixels are considered as an abnormal example of the background pixels, and the removal method of the anomaly samples is proposed. Then the fuzzy estimation operator is trained by the removal results. The method has high efficiency and detection in the dynamic background, the illumination change, the camera vibration and so on. Effect.
2. a motion shadow detection method based on fuzzy integral is proposed. On the basis of extracting foreground region, color and texture are selected as the feature of shadow detection, and the similarity and importance measure function of the two features are defined respectively. Then the two features are fused by Choquet fuzzy integral to realize the shadow and foreground object. Classification, and finally through the subsequent processing, find the real shadow area.
3. a target classification method based on JointBoost I2C distance measurement is proposed. In view of the shortage of classical I2C distance computation and easy to be disturbed by noise interference, a new method of generating prototype feature sets is proposed. The number of samples in the set is less, but more representative, the distance cost of the test image to the prototype feature set is calculated. Less time; then the idea of JointBoost algorithm is used to combine multiple I2C distance metrics to generate a strong classifier. Finally, a method of fusion of spatial information to a strong classifier is proposed. Experiments show that the method has a higher classification performance in the foreground object and the image classification experiment.
4. a target recognition method based on characteristic codebook tree and energy minimization is proposed. This method takes into account the spatial location information of the feature and the spatial relationship between features, and integrates target detection and target recognition. First, the noise features are filtered out from the large number of features extracted from the target image, and then the single feature and the adjacent space in the space are connected in series. The double feature uses the hierarchical K mean clustering algorithm to construct the characteristic tree tree. The tree model can be used to locate and classify the features quickly. Finally, an energy function is established to fuse the probability matching results of the single, double feature codebook, and to find the energy minimization of the region in which the sliding window is located in the test image. The position of the category target.
5. the target recognition method based on optimized Hough forest cost loss is proposed. Firstly, based on the known location of the object in the training image, an improved measurement method of offset uncertainty is proposed. Secondly, the adaptive weight distribution of the sample of picture block and target object is learned with the help of the thought of the Boosting algorithm. The cost loss functions used to construct random trees and Hough forests are optimized respectively. Finally, based on the weight distribution of the block samples, an improved method for measuring the uncertainty of the class marks is proposed. Based on the cost loss function of the Hough forest, the learning method of the weight of the random tree is also proposed.
【学位授予单位】：华南理工大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：U495

【参考文献】