Showing papers by "Simon Maskell published in 2006"

Fast particle smoothing: if I had a million particles

Abstract: We propose efficient particle smoothing methods for generalized state-spaces models. Particle smoothing is an expensive O(N2) algorithm, where N is the number of particles. We overcome this problem by integrating dual tree recursions and fast multipole techniques with forward-backward smoothers, a new generalized two-filter smoother and a maximum a posteriori (MAP) smoother. Our experiments show that these improvements can substantially increase the practicality of particle smoothing.

Real-Time Tracking Of Hundreds Of Targets With Efficient Exact JPDAF Implementation

Abstract: An assignment problem is considered with the constraint that the same hypothesis cannot be applied to more than one object. We desire efficiency without approximation. Multiple target tracking methods such as the Joint Probabilistic Association Filter (JPDAF) motivate us. Methods of solving this assignment problem involving enumerating all possible joint assignments are infeasible except for small problems. A recent approach circumvents this combinatorial explosion by representing the structure of the target hypotheses in a `net' which exploits redundancy in an ordered list of objects used to describe the problem. Here, we generalize this approach to process the objects in a tree structure; this exploits conditional independence between subsets of the objects. This gives a substantial computational saving and allows us to consider scenarios which were previously impractical. In particular, we show the feasibility of using an exact JPDAF implementation to track 400 targets.

Joint Tracking and Classification of Airbourne Objects using Particle Filters and the Continuous Transferable Belief Model

Abstract: This paper describes the integration of a particle filter and a continuous version o f the transferable belief model. The output from the particle filter is used as input to the transferable belief model. The transferable belief model's continuous nature allows for the prior knowledge over the classification space to be incorporated within the system. Classification of objects is demonstrated within the paper and compared to the more classical Bayesian classification routine. This is the first time that such an approach has been taken to jointly classify and track targets. We show that there is a great deal o f flexibility built into the continuous transferable belief model and in our comparison with a Bayesian classifier, we show that our novel approach offers a more robust classification output that is less influenced by noise.

A Single Instruction Multiple Data Particle Filter

Abstract: Particle filters are often claimed to be readily parallelisable However, the resampling step is non-trivial to implement in a fine-grained parallel architecture While approaches have been proposed that modify the particle filter to be amenable to such implementation, this paper's novelty lies in its description of a Single Instruction Multiple Data (SIMD) implementation of a particle filter that uses N processors to process N particles The resulting algorithm has a time complexity of O((log N)2) when performing resampling using N processors The algorithm has been implemented using C for Graphics (CG), a language that enables the heavily pipelined architecture of modern graphics cards to be used to imitate a SIMD processor Initial results are presented

Distributed Tracking of Stealthy Targets using Particle Filters

Abstract: This paper presents a novel distributed particle filter algorithm. To solve the problem of fusing the output of multiple particle filters, a joint space over multiple realisations of the same variable is used. This approach to using particle filters to perform distributed tracking of stealthy targets requires minimal modifications to the particle filters running at the sensor nodes and does not necessitate data to be transmitted to the fusion node.

Fixed-lag sequential Monte Carlo data association

Abstract: The use of multiple scans of data to improve ones ability to improve target tracking performance is widespread in the tracking literature. In this paper, we introduce a novel application of a recent innovation in the SMC literature that uses multiple scans of data to improve the stochastic approximation (and so the data association ability) of a multiple target Sequential Monte Carlo based tracking system. Such an improvement is achieved by resimulating sampled variates over a fixed-lag time window by artificially extending the space of the target distribution. In doing so, the stochastic approximation is improved and so the data association ambiguity is more readily resolved.