Data-Driven Approaches for Modelling Target Behaviour

TL;DR

A comparative study between three different methods that use machine learning to describe the underlying object motion based on training data and their respective strengths are highlighted in one simulated and two real-world scenarios.

Abstract

The performance of tracking algorithms strongly depends on the chosen model assumptions regarding the target dynamics. If there is a strong mismatch between the chosen model and the true object motion, the track quality may be poor or the track is easily lost. Still, the true dynamics might not be known a priori or it is too complex to be expressed in a tractable mathematical formulation. This paper provides a comparative study between three different methods that use machine learning to describe the underlying object motion based on training data. The first method builds on Gaussian Processes (GPs) for predicting the object motion, the second learns the parameters of an Interacting Multiple Model (IMM) filter and the third uses a Long Short-Term Memory (LSTM) network as a motion model. All methods are compared against an Extended Kalman Filter (EKF) with an analytic motion model as a benchmark and their respective strengths are highlighted in one simulated and two real-world scenarios.

Figures (9)

• Figure 1: The factor graph of the proposed joint GP algorithm for state prediction mainly includes offline training and online prediction processes. Legend: Circles – variable nodes; Squares – factor nodes. The process for the prediction on the X-axis is marked in blue.
• Figure 2: The lstm network architecture used for the mkf implementation. The arrows display the number of outputs of the respective layer: $d_x$ is the dimension of the target state space, $n_\text{HU}^L$ and $n_\text{HU}^D$ denote the number of hidden units of the lstm and dense layers, respectively, and $d_N=0.5 d_x (d_x+3)$.
• Figure 3: Sample trajectory (---) and synthetic measurements ($\times$), $100$ time steps. The simulated sensor is located at the origin.
• Figure 4: uav dataset, relative east/north positions to the start position over the whole flight duration. The hypothetical observer ($\blacktriangle$) is placed at $[25m,-50m]$ relative to the start position $[0,0]$.
• Figure 5: rib datasets recorded on two different days of the trial. The assumed sensor ($\blacktriangle$) is placed at the Leonardo headquarters, which is about 11km east and 2.8km south of Port Edgar where the rib started on both days.