基于神经网络与多特征融合的粒子滤波目标检测跟踪算法研究

发布时间：2018-07-26 13:20

【摘要】：运动目标的检测和跟踪作为数字图像处理和计算机视觉领域的热点,在自动导航、交通监控、国防军工等领域都具有十分高的应用价值。过去几十年,众多研究者在目标检测和跟踪领域进行了深入的研究,但是由于应用场景复杂多变、目标运动规律复杂等因素,导致目标的检测与跟踪技术无法得到大量广泛的使用。因此,设计一种适用性强的目标检测与跟踪算法有着重大的意义。针对目标跟踪检测的有效性与准确性,本文在融合帧差法的混合高斯模型目标区域提取基础上,研究基于多特征的粒子滤波目标跟踪算法,并且利用BP神经网络调整与改进粒子滤波跟踪算法。论文主要的算法改进和成果如下:一、提出一种融合帧差法的混合高斯模型获取运动目标区域。混合高斯模型不能完整的检测运动前景区域,容易将背景错误混淆成前景。本文通过融合帧差法和混合高斯模型,通过区分背景凸显区域与前景区域,利用不同的学习速率来完整的提取目标运动区域。二、提出了基于多种特征融合的粒子滤波跟踪方法。针对单个特征的跟踪模型算法精度低,适用性差等问题。本文提取目标的颜色特征和HOG特征,构建多特征观测模型,通过该特征模型进行粒子滤波目标检测与跟踪。实验表明该算法准确性更高。三、提出了一种通过BP神经网络改进多特征融合的粒子滤波跟踪算法。传统的粒子滤波算法具有粒子退化问题,粒子数越来越匮乏。本文利用BP神经网络反向传播来调整更新粒子权值,增加粒子的多样性,通过结合多特征模型,改善算法的滤波性能,并且提高了目标跟踪的精度。本文通过以上三个方面对粒子滤波的目标检测跟踪算法做了优化改进,实验结果表明,在复杂的场景如发生遮挡、背景相似、运动规律多样复杂等情况下,目标跟踪的误差得到缩减,精度获得了相应的提升。
[Abstract]:As a hot spot in the field of digital image processing and computer vision, the detection and tracking of moving targets has high application value in the fields of automatic navigation, traffic monitoring, national defense industry and so on. In the past few decades, many researchers have carried out in-depth research in the field of target detection and tracking. As a result, target detection and tracking technology can not be widely used. Therefore, it is of great significance to design a target detection and tracking algorithm with strong applicability. Aiming at the validity and accuracy of target tracking detection, this paper studies the particle filter target tracking algorithm based on multi-feature based on the target region extraction of hybrid Gao Si model based on fusion frame difference method. And the BP neural network is used to adjust and improve the particle filter tracking algorithm. The main improvements and results of this paper are as follows: firstly, a hybrid Gao Si model based on frame difference method is proposed to obtain the moving target region. The mixed Gao Si model can not detect the moving foreground region completely, so it is easy to confuse the background error with the foreground. By combining frame difference method and hybrid Gao Si model, this paper distinguishes the background salient region from the foreground region, and uses different learning rates to extract the moving region of the target completely. Secondly, a particle filter tracking method based on multi-feature fusion is proposed. The tracking model based on single feature has low accuracy and poor applicability. In this paper, the color features and HOG features of the target are extracted, and a multi-feature observation model is constructed, which is used for particle filter target detection and tracking. Experiments show that the algorithm is more accurate. Thirdly, an improved particle filter tracking algorithm based on BP neural network is proposed. The traditional particle filter algorithm has the problem of particle degradation, and the number of particles is becoming more and more scarce. In this paper, BP neural network is used to adjust the weight of the updated particle, to increase the diversity of particles, to improve the filtering performance of the algorithm and to improve the precision of target tracking by combining the multi-feature model. In this paper, the target detection and tracking algorithm of particle filter is optimized and improved through the above three aspects. The experimental results show that, in the case of complex scenes such as occlusion, similar background, complex motion rules and so on, The target tracking error is reduced and the accuracy is improved accordingly.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】