The challenges facing unmanned underwater vehicle (UUV) target state estimation are observation uncertainty, the unpredictability of the target motion model, and the complex relative motion of the moving target to the observer. Aiming at the above problems, a convolutional neural network particle filter (CNNPF) is proposed and applied to UUV target state estimation using forward-looking sonar. First, we design a target state prediction network based on convolutional neural network (CNN) to describe the nonlinear measurement and non-Markov target motion process models. Then, a target state prediction dataset is established for the prediction network to build the state space of UUV target state estimation and extract target motion features from the nonsequential measurements. Moreover, particles are sampled from the observations with disturbed time series to approximate the non-Gaussian distribution of measurements. The sampling particles in the target measured state space are mapped to target predictive state space through the prediction network. Convolutional and pooling layers impart invariance to the prediction network and improve its adaptability to uncertain measurement. Finally, the predicted states of particles are used to estimate the target states. The statistical experiment, simulation cases, and hardware-in-the-loop experiment proved the applicability and performance of the CNNPF.

中文翻译：

---

一种用于 UUV 目标状态估计的卷积神经网络粒子滤波器

无人水下航行器（UUV）目标状态估计面临的挑战是观测不确定性、目标运动模型的不可预测性以及运动目标与观察者的复杂相对运动。针对上述问题，提出了一种卷积神经网络粒子滤波器（CNNPF），并将其应用于使用前视声纳的UUV目标状态估计。首先，我们设计了一个基于卷积神经网络（CNN）的目标状态预测网络来描述非线性测量和非马尔可夫目标运动过程模型。然后，为预测网络建立目标状态预测数据集，构建 UUV 目标状态估计的状态空间，并从非序列测量中提取目标运动特征。而且，从具有扰动时间序列的观察中对粒子进行采样，以近似测量的非高斯分布。目标测量状态空间中的采样粒子通过预测网络映射到目标预测状态空间。卷积层和池化层赋予预测网络不变性并提高其对不确定测量的适应性。最后，使用粒子的预测状态来估计目标状态。统计实验、仿真案例和硬件在环实验证明了CNNPF的适用性和性能。卷积层和池化层赋予预测网络不变性并提高其对不确定测量的适应性。最后，使用粒子的预测状态来估计目标状态。统计实验、仿真案例和硬件在环实验证明了CNNPF的适用性和性能。卷积层和池化层赋予预测网络不变性并提高其对不确定测量的适应性。最后，使用粒子的预测状态来估计目标状态。统计实验、仿真案例和硬件在环实验证明了CNNPF的适用性和性能。