A cross-and-dot-product neural network based filtering for maneuvering-target tracking

Digital Twins Based Intelligent State Prediction Method for Maneuvering-Target Tracking

“…Once the target states or observation noises are out of the ranges of training data, performances of those data-driven methods will severely degrade. To avoid training the network to adapt to such data with widely varying range, Liu [19] further proposed a cross-and-dot-product neural network (CADP-NN) to learn the motion-models, which vary relatively less. Thus, the prediction performance of CADP-NN becomes more stable when inputs vary out of the training sets.…”

“…Hence, CADP-NN needs common UKF with predefined motion-models to offer predicted states. Although there is a double-channels method in [19] to maximize the use of estimated motion-models from CADP-NN to improve tracking performance, the actual tracking accuracy is still limited by the UKF itself. Thus, the tracking performance of CADP-NN is still not good enough.…”

“…In (2), we know that 𝑭 is related to target states. This relationship has been deduced in [19] by the CADP-NN. But in fact, target states contain both position and velocity information, yet observation only contains the information of position without velocity.…”

“…To this end, two tracking scenes are designed, which are the low-maneuvering and highmaneuvering motion scenes, respectively. Then, in comparison with the state-of-the-art data-driven methods: DeepMTT [15] and CADP-NN [19] and model-driven method: MIE-BLUE-IMM [14], we discuss the tracking performance of our ISPM based on the tracking RMSE defined as follows,…”

“…In the tracking process, our ISPM method consumes 17.8 ms to compute one iteration in tracking process. According to[15],[19], DeepMTT, CADP-NN and MIE-BLUE-IMM methods consume 16.4 ms, 2.0ms and 0.6 ms, respectively. Although the computational time of our ISPM for one iteration of tracking is the longest in comparison with other methods, 17.8ms per iteration is in fact short enough for practical…”