In order to improve the performance of the probability hypothesis density(PHD) algorithm based particle filter(PF) in terms of number estimation and states extraction of multiple targets, a new probability hypothesis density filter algorithm based on marginalized particle and kernel density estimation is proposed, which utilizes the idea of marginalized particle filter to enhance the estimating performance of the PHD. The state variables are decomposed into linear and non-linear parts. The particle filter is adopted to predict and estimate the nonlinear states of multi-target after dimensionality reduction, while the Kalman filter is applied to estimate the linear parts under linear Gaussian condition. Embedding the information of the linear states into the estimated nonlinear states helps to reduce the estimating variance and improve the accuracy of target number estimation. The meanshift kernel density estimation, being of the inherent nature of searching peak value via an adaptive gradient ascent iteration, is introduced to cluster particles and extract target states, which is independent of the target number and can converge to the local peak position of the PHD distribution while avoiding the errors due to the inaccuracy in modeling and parameters estimation. Experiments show that the proposed algorithm can obtain higher tracking accuracy when using fewer sampling particles and is of lower computational complexity compared with the PF-PHD.
A fast algorithm based on the grayscale distribution of infrared target and the weighted kernel function was proposed for the moving target detection(MTD) in dynamic scene of image series. This algorithm is used to deal with issues like the large computational complexity, the fluctuation of grayscale, and the noise in infrared images. Four characteristic points were selected by analyzing the grayscale distribution in infrared image, of which the series was quickly matched with an affine transformation model. The image was then divided into 32×32 squares and the gray-weighted kernel(GWK) for each square was calculated. At last, the MTD was carried out according to the variation of the four GWKs. The results indicate that the MTD can be achieved in real time using the algorithm with the fluctuations of grayscale and noise can be effectively suppressed. The detection probability is greater than 90% with the false alarm rate lower than 5% when the calculation time is less than 40 ms.
目的在显著物体检测算法发展过程中,基准数据集发挥了重要作用。然而,现有基准数据集普遍存在数据集偏差,难以充分体现不同算法的性能,不能完全反映某些典型应用的技术特点。针对这一问题,本文对基准数据集的偏差和统计特性展开定量分析,提出针对特定任务的新基准数据集。方法首先,讨论设计和评价数据集时常用的度量指标;然后,定量分析基准数据集的统计学差异,设计新的基准数据集MTMS300(multiple targets and multiple scales);接着,使用基准数据集对典型视觉显著性算法展开性能评估;最后,从公开基准数据集中找出对多数非深度学习算法而言都较为困难(指标得分低)的图像,构成另一个基准数据集DSC(difficult scenes in common)。结果采用平均注释图、超像素数目等6种度量指标对11个基准数据集进行定量分析,MTMS300数据集具有中心偏差小、目标面积比分布均衡、图像分辨率多样和目标数量较多等特点,DSC数据集具有前景/背景差异小、超像素数量多和图像熵值高的特点。使用11个基准数据集对18种视觉显著性算法进行定量评估,揭示了算法得分和数据集复杂度之间的相关性,并在MTMS300数据集上发现了现有算法的不足。结论提出的两个基准数据集具有不同的特点,有助于更为全面地评估视觉显著性算法,推动视觉显著性算法向特定任务方向发展。
