Probabilistic models for sequential data are the basis for a variety of applications concerned with processing timely ordered information. The predominant approach in this domain is given by neural networks, which incorporate either stochastic units or components. This paper proposes a new probabilistic sequence model building on probabilistic B\'ezier curves. Using Gaussian distributed control points, these parametric curves pose a special case for Gaussian processes (GP). Combined with a Mixture Density network, Bayesian conditional inference can be performed without the need for mean field variational approximation or Monte Carlo simulation, which is a requirement of common approaches. For assessing this hybrid model's viability, it is applied to an exemplary sequence prediction task. In this case the model is used for pedestrian trajectory prediction, where a generated prediction also serves as a GP prior. Following this, the initial prediction can be refined using the GP framework by calculating different posterior distributions, in order to adapt more towards a given observed trajectory segment.
翻译:连续数据概率模型是处理及时定购信息的各种应用的基础。这一领域的主要方法由神经网络提供,这些网络包含随机单位或组件。本文件提议在概率B\'ezier曲线的基础上建立一个新的概率序列模型。使用高西亚分布式控制点,这些参数曲线为高山进程提供了一个特殊案例。与混合代号网络相结合,贝耶斯有条件的推断可以进行,而不需要平均场外变化近似值或蒙特卡洛模拟,这是共同方法的一项要求。为了评估这种混合模型的可行性,该模型应用到模拟序列预测任务中。在这种情况下,模型用于行人轨迹预测,而生成的预测也在此之前也起到GP的作用。在此之后,可以利用GP框架来计算不同的远地点分布来改进初步预测,以便更适应特定的观察到的轨迹段。