DeepGTT: A general trajectory tracking deep learning algorithm based on dynamic law learning

Abstract

Deep learning technology provides novel solutions for increasingly complex target tracking requirements. For traditional target tracking models, the movement of the target need to be simulated by a predefined mathematical model. However, it is extremely difficult to obtain sufficient information in advance, which makes it challenging to track changeable and noisy trajectories in a timely and precise manner. A deep learning framework is constructed for automatic trajectory tracking based on learning the dynamic laws of motion, called DeepGTT. Specifically, a trajectory generator and a trajectory mapper were designed to standardise trajectory data and construct trajectory mapping, which enable the long short-term memory-based tracking network to learn general dynamic laws. Then, to discuss the interpretability of the model, the mechanism of the deep learning framework is considered and a memory factor matrix is defined. Finally, extensive experiments are conducted on various weak manoeuvring and manoeuvring scenarios to evaluate the algorithm. Experimental results demonstrate that the DeepGTT algorithm remarkably improves accuracy and efficiency compared with most conventional algorithms and state-of-the-art methods. In addition, interpretability experiments qualitatively prove that the tracking network can perceive dynamic laws when estimating the target state.This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.

A deep learning algorithm for joint direct tracking and classification of manoeuvring sources

“…To verify the feasibility of our algorithm, we employ the dynamic model to generate two types of targets. We use a set of models containing the constant velocity (CV), constant turn (CT), and constant turn rate velocity (CTRV) to describe the manoeuvring motion [32]. The typical parameters of class A and class B are shown in Table 2.…”

“…Although the aircraft follow similar dynamic laws, different types of aircraft, such as helicopters and transport aeroplanes, are always diverse in their speed range, state noise, and manoeuvering characteristics. Typical dynamic models can be found in [32]. Specifically, the Equation (1) in CV model is where is the discrete‐time state of the trajectory, is the transition noise sequence and is usually modelled as Gaussian noise.…”

“…Although the aircraft follow similar dynamic laws, different types of aircraft, such as helicopters and transport aeroplanes, are always diverse in their speed range, state noise, and manoeuvering characteristics. Typical dynamic models can be found in [32]. Specifically, the Equation (1) in CV model is…”

“…Yu et al [32] constructed an LSTM-based model by learning dynamic laws, and realised real-time and accurate tracking of various manoeuvering motions. The interpretability experiments in [32] further indicate that dynamic laws of the manoeuvering target are captured by the network during tracking. Nevertheless, existing tracking networks only consider the two-step tracking, and mainly utilise RNNs, such as LSTM and gated recurrent unit (GRU) [33].…”

“…Yu et al. [32] constructed an LSTM‐based model by learning dynamic laws, and realised real‐time and accurate tracking of various manoeuvering motions. The interpretability experiments in [32] further indicate that dynamic laws of the manoeuvering target are captured by the network during tracking.…”