Sequential Monte Carlo Filtering with Long Short-Term Memory Prediction

Recursive Bayesian estimation builds on the use of suitable mathematical models in order to accurately describe the evolution of an object based on imperfect sensor data. Sequential Monte Carlo (SMC) techniques have been developed to overcome the need of restrictive models, however more sophisticated motion models can require high-dimensional particle states, resulting in a need for many particles and hence greater computation. Additionally, most conventional prediction methods assume low-order Markovian transitions and therefore tend to ignore the majority of the target history that may be relevant to the prediction. In this article, a variation of the Particle Filter (PF) is introduced which uses a Long Short-Term Memory (LSTM) neural network to perform the prediction step. In contrast to existing techniques, the network is trained to predict the relative displacement between object states rather than absolute positions to gain rotation and translation invariance. Experiments on simulated trajectories show that the approach is more robust against low probability of detection and occlusion than the conventional PF.