Syntactic Enhancement to VSIMM for Roadmap Based Anomalous Trajectory Detection: A Natural Language Processing Approach

Abstract

Syntactic tracking aims to classify a target's spatiotemporal trajectory by using natural language processing models. This paper proposes constrained stochastic context free grammar (CSCFG) models for target trajectories confined to a roadmap. We present a particle filtering algorithm that exploits the CSCFG model structure to estimate the target's trajectory. This metalevel algorithm operates in conjunction with a base-level target tracking algorithm. Extensive numerical results using simulated ground moving target indicator (GMTI) radar measurements show useful improvement in both trajectory classification and target state (both coordinates and velocity) estimation.Keywords-syntactic tracking, constrained stochastic context free grammar (CSCFG), particle filter, Earley Stolcke parser 1 Traditionally, given track information, a radar (human) operator examines target trajectories to determine anomalous behavior. This paper develops natural language models and meta-level signal processing algorithms for estimating anomalous trajectories. Such "middleware" forms the interface between the physical signal processing layer and the radar operator; see also [7] for alternative models for intent using bridging distributions.

On Destination Prediction Based on Markov Bridging Distributions

“…Algorithm 1 (mean run-time is 2.62ms at each t k ) shows a reduction of around 65% compared to the original BD (mean 7.42ms); Algorithm 2 has mean run-time of 5.36ms. This confirms the complexity analysis in (14)- (16) with parameters {m, s, q, N } = {2, 4, 15, 6} and demonstrates the potential of the proposed efficient methods for real-time implementations.…”

“…Additionally, OU processes, with a priori learnt means, are shown to reliably model and estimate vessels motion in maritime surveillance [7], [8], [9]. Discretised state-space models based on reciprocal processes or other models from natural language processing are proposed in [15], [16] to recognise the object intent. They stipulate that the target should pass through a finite number of predefined spatial grid cells to reach its endpoint.…”

“…They stipulate that the target should pass through a finite number of predefined spatial grid cells to reach its endpoint. The proposed approach in this letter aims to predict the unknown destination D, not estimate x n , and has notably lower complexity compared with those in [15], [16]. It utilises continuous-time state space models to treat asynchronous measurements and capture, via a Markov bridge, the long term underlying dependencies in the object trajectory as dictated by the intended endpoint.…”

“…Table I depicts the complexity order of all methods as a function of the state dimension s; complexity order for m or any other parameter can be similarly attained from (15) and (16). As well as (14)- (16), the table clearly illustrates that the proposed formulation is significantly more efficient compared with the original BD. Algorithm 1 can reduce the computational complexity of the destination inference routine by approximately 57.5%; 65% for motion models independent of D i .…”

Bayesian Intent Prediction for Fast Maneuvering Objects using Variable Rate Particle Filters

“…Intent inference methods that utilise stochastic context-free grammars and reciprocal process were proposed in [23,24] and extended to apply to nonlinear systems with particle-filter-based schemes in [1] and [25]. The underlying premise of these techniques is that the tracked object can follow a set of predefined trajectories within a discritised state space formulation.…”