Parallel Resampling Research Articles

High-Resolution Feature Pyramid Network for Small Object Detection on Drone View

Object detection has developed rapidly with the help of deep learning technologies recent years. However, object detection on drone view remains challenging due to two main reasons: (1) It is difficult to detect small-scale objects lacking detailed information. (2) The diversity of camera angles of drones brings dramatic differences in object scale. Although feature pyramid network (FPN) alleviates the problem caused by scale difference to some extent, it also retains some worthless features, which wastes resources and slows down the speed. In this work, we propose a novel High-Resolution Feature Pyramid Network (HR-FPN) to improve the detection accuracy of small-scale objects and avoid feature redundancy. The key components of HR-FPN include a high-resolution feature alignment module (HRFA), a high-resolution feature fusion module (HRFF) and a multi-scale decoupled head (MSDH). HRFA feeds multi-scale features from backbone into parallel resampling channels to obtain high-resolution features at the same scale. HRFF establishes a bottom-up path to distribute context-rich low-level semantic information to all layers that are then aggregated into classification feature and localization feature. MSDH cope with the scale difference of objects by predicting the categories and locations corresponding to different scales of objects separately. Moreover, we train model by scale-weighted loss to focus more on small-scale objects. Extensive experiments and comprehensive evaluations demonstrate the effectiveness and advancement of our approach.

High-Throughput Aqueous Electrolyte Structure Prediction Using IonSolvR and Equivariant Graph Neural Network Potentials.

Neural network potentials have recently emerged as an efficient and accurate tool for accelerating ab initio molecular dynamics (AIMD) in order to simulate complex condensed phases such as electrolyte solutions. Their principal limitation, however, is their requirement for sufficiently large and accurate training sets, which are often composed of Kohn-Sham density functional theory (DFT) calculations. Here we examine the feasibility of using existing density functional tight-binding (DFTB) molecular dynamics trajectory data available in the IonSolvR database in order to accelerate the training of E(3)-equivariant graph neural network potentials. We show that the solvation structure of Na+ and Cl- in aqueous NaCl solutions can be accurately reproduced with remarkably small amounts of data (i.e., 100 MD frames). We further show that these predictions can be systematically improved further via an embarrassingly parallel resampling approach.

Efficient Massive Computing for Deformable Volume Data Using Revised Parallel Resampling

In this paper, we propose an improved parallel resampling technique. Parallel resampling is a deformable object generation method based on volume data applied to medical simulations. Existing parallel resampling is not suitable for massive computing, because the number of samplings is high and floating-point precision problems may occur. This study addresses these problems to obtain improved user latency when performing medical simulations. Specifically, instead of interpolating values after volume sampling, the efficiency is improved by performing volume sampling after coordinate interpolation. Next, the floating-point error in the calculation of the sampling position is described, and the advantage of barycentric interpolation using a reference point is discussed. The experimental results showed a significant improvement over the existing method. Volume data comprising more than 600 images used in clinical practice were deformed and rendered at interactive speed. In an Internet of Everything environment, medical imaging systems are an important application, and simulation image generation is also valuable in the overall system. Through the proposed method, the performance of the whole system can be improved.

FPGA Implementation of Particle Filters for Robotic Source Localization

Particle filtering is very reliable in modelling non-Gaussian and non-linear elements of physical systems, which makes it ideal for tracking and localization applications. However, a major drawback of particle filters is their computational complexity, which inhibits their use in real-time applications with conventional CPU or DSP based implementation schemes. The re-sampling step in the particle filters creates a computational bottleneck since it is inherently sequential and cannot be parallelized. This paper proposes a modification to the existing particle filter algorithm, which enables parallel re-sampling and reduces the effect of the re-sampling bottleneck. We then present a high-speed and dedicated hardware architecture incorporating pipe-lining and parallelization design strategies to supplement the modified algorithm and lower the execution time considerably. From an application standpoint, we propose a novel source localization model to estimate the position of a source in a noisy environment using the particle filter algorithm implemented on hardware. The design has been prototyped using Artix-7 field-programmable gate array (FPGA), and resource utilization for the proposed system is presented. Further, we show the execution time and estimation accuracy of the high-speed architecture and observe a significant reduction in computational time. Our implementation of particle filters on FPGA is scalable and modular, with a low execution time of about 5.62 μs for processing 1024 particles (compared to 64 ms on Intel Core i7-7700 CPU with eight cores clocking at 3.60 GHz) and can be deployed for real-time applications.

Improved Parallel Resampling Methods for Particle Filtering

Particle filter techniques are common methods used to estimate the evolving state of nonlinear, non-Gaussian time-variant systems by utilizing a periodic sequence of noisy measurements. The accuracy of particle filter methods has often been shown to be superior to other state estimation techniques, such as the extended Kalman filter (EKF), for many applications. Unfortunately, the high computational cost and highly nondeterministic runtime behavior of particle filters often preclude their use in hard, real-time environments, where filter response must meet the strict timing requirements of the application. Particle filter algorithms are composed of three main stages: prediction, update, and resampling. General purpose graphics processing units (GPGPUs) have been successfully employed in previous research to accelerate the computation of both the prediction and update stages by exploiting their natural fine-grain parallelism. This research focuses on accelerating the resampling stage for (GPGPU) execution, which has been much more difficult to parallelize due to it's apparent inherent sequentially. This paper introduces a novel GPGPU implementation of the systematic and stratified resampling algorithms that exploit the monotonically increasing nature of the prefix-sum and the evolutionary nature of the particle weighting process to allow the re-indexing portion of the algorithms to occur in a two-phase, multi-threaded manner. This resulting measured factor of performance improvement for the systematic and stratified algorithms was 15$\times$ and 32$\times$ , respectively, over the serial implementations.

Parallel Resampling of OFDM Signals for Fluctuating Doppler Shifts in Underwater Acoustic Communication

Reliable underwater acoustic communication is demanded for autonomous underwater vehicles (AUVs) and remotely operated underwater vehicles (ROVs). Orthogonal frequency-division multiplexing (OFDM) is robust with multipath interference; however, it is sensitive to Doppler. Doppler compensation is given by two-step processing of resampling and residual carrier frequency offset (CFO) compensation. This paper describes the improvement of a resampling technique. The conventional method assumes a constant Doppler shift during a communication frame. It cannot cope with Doppler fluctuation, where relative speeds between transmitter and receiver units are fluctuating. We propose a parallel resampling technique that a resampling range is extended by measured Doppler standard deviation. The effectiveness of parallel resampling has been confirmed in the communication experiment. The proposed method shows better performance in bit error rates (BERs) and frame error rates (FERs) compared with the conventional method.

Optimization of Particle CBMeMBer Filters for Hardware Implementation

It is a promising solution for real-time multitarget tracking to implement particle cardinality balanced multitarget multi-Bernoulli (CBMeMBer) filters in hardware platforms. However, this solution is difficult to materialize since there is a contradiction between the time-varying number of Bernoulli intensity components in CBMeMBer filters and the limited number of particles in hardware platforms. Moreover, real-time hardware implementation requires a resampling procedure that is suitable for parallel processing, while the existing parallel resampling algorithms oversimplify this procedure, resulting in estimation performance degradation. In this paper, we propose an optimization algorithm of particle allocation to overcome the above-mentioned contradiction, and a parallel resampling algorithm to improve the estimation performance. Numerical experiments demonstrate the effectiveness of the proposed algorithms in multitarget tracking.

Deep-pipelined FPGA Implementation of Real-time Object Tracking using a Particle Filter

This paper presents a real-time FPGA implementation of the posterior system state estimation in dynamic models, which is developed using particle filter algorithm. Specially, our system is constructed by parallel resampling (FO-resampling) algorithm on a stream-based architecture. To be precise, the resampling is accomplished in a valid pixel area of an input image frame while prediction and update of particles are performed in a synchronization region, thus our approach achieves realtime performance of 60 fps for VGA images, synchronized with the camera pixel throughput without using any external memory devices. Through evaluation with an object tracking benchmark video, the tradeoff relationship between tracking quality and the number of particles is analyzed to find an appropriate hardware parameters. In addition, we address improvement of resource utilization for our particle filter architecture, especially by using a higher clock frequency to reuse hardware resources in a time sharing manner. The implementation experiments reveal that the proposed approach allows the original design to be fitted in a smaller FPGA chip. However, we also demonstrate this size reduction approach has an overhead of 2.7 to 3.0 times power consumption compared to original designs with a slow clock frequency.

Parallel Resampling in the Particle Filter

Modern parallel computing devices, such as the graphics processing unit (GPU), have gained significant traction in scientific and statistical computing. They are particularly well-suited to data-parallel algorithms such as the particle filter, or more generally sequential Monte Carlo (SMC), which are increasingly used in statistical inference. SMC methods carry a set of weighted particles through repeated propagation, weighting, and resampling steps. The propagation and weighting steps are straightforward to parallelize, as they require only independent operations on each particle. The resampling step is more difficult, as standard schemes require a collective operation, such as a sum, across particle weights. Focusing on this resampling step, we analyze two alternative schemes that do not involve a collective operation (Metropolis and rejection resamplers), and compare them to standard schemes (multinomial, stratified, and systematic resamplers). We find that, in certain circumstances, the alternative resamplers can perform significantly faster on a GPU, and to a lesser extent on a CPU, than the standard approaches. Moreover, in single precision, the standard approaches are numerically biased for upward of hundreds of thousands of particles, while the alternatives are not. This is particularly important given greater single- than double-precision throughput on modern devices, and the consequent temptation to use single precision with a greater number of particles. Finally, we provide auxiliary functions useful for implementation, such as for the permutation of ancestry vectors to enable in-place propagation. Supplementary materials are available online.

Parallel particle PHD filter implemented on multicore and cluster systems

The Probability Hypothesis Density (PHD) filter is a promising technique in terms of computational complexity to solve the multiple targets tracking problem. However, the amount of computation is prohibitive in critical situations when the clutter intensity and sample rate are high. Therefore, the execution time of the sequential particle PHD filter cannot meet the requirement for real-time processing applications. To address this problem, we propose a parallel scheme for efficient implementation of particle PHD filter on clusters of multicore distributed memory architecture. Since particles can be treated separately and spread among processors, the prediction and update step can be readily performed in parallel. However, the resampling and estimation step become the bottleneck that significantly affects the speedup and scalability achieved by the parallel implementation of particle PHD filter for the requirement of joint processing of all particles. We propose an approach to fulfill parallel resampling and stratified estimation in a unified architecture. Particle exchange to rebalance the work load among computing nodes is also discussed. Experiment results show that tracking performance of the parallel version is almost equivalent to or even better than the sequential one, while in terms of execution time we can achieve a tremendous speedup.

A parallel resampling method for interactive deformation of volumetric models

In this work, we propose a method to interactively deform high-resolution volumetric datasets, such as those obtained through medical imaging. Interactive deformation enables the visualization of these datasets in full detail using state-of-the-art volume rendering techniques as they are dynamically modified. Our approach relies on resampling the original dataset to a target regular grid, following a 3D rasterization technique. We employ an implicit auxiliary mesh to execute resampling, which allows us to decouple mapping of the deformation field to the volume from actual resampling. In this way, our method is practically independent of the deformation method of choice, as well as of the resolution of the deformation meshes. We show how our method lends itself nicely to an efficient, massively parallel implementation on GPUs, and we demonstrate its application on several high-resolution datasets and deformation models.

Particle filter based on iterated importance density function and parallel resampling

The design, analysis and parallel implementation of particle filter (PF) were investigated. Firstly, to tackle the particle degeneracy problem in the PF, an iterated importance density function (IIDF) was proposed, where a new term associating with the current measurement information (CMI) was introduced into the expression of the sampled particles. Through the repeated use of the least squares estimate, the CMI can be integrated into the sampling stage in an iterative manner, conducing to the greatly improved sampling quality. By running the IIDF, an iterated PF (IPF) can be obtained. Subsequently, a parallel resampling (PR) was proposed for the purpose of parallel implementation of IPF, whose main idea was the same as systematic resampling (SR) but performed differently. The PR directly used the integral part of the product of the particle weight and particle number as the number of times that a particle was replicated, and it simultaneously eliminated the particles with the smallest weights, which are the two key differences from the SR. The detailed implementation procedures on the graphics processing unit of IPF based on the PR were presented at last. The performance of the IPF, PR and their parallel implementations are illustrated via one-dimensional numerical simulation and practical application of passive radar target tracking.

A Parallel Systematic Resampling Algorithm for High-Speed Particle Filters in Embedded Systems

In this paper, we propose a parallel systematic resampling (PSR) algorithm for particle filters, which is a new form of systematic resampling (SR). The PSR algorithm makes iterations independent, thus allowing the resampling algorithm to perform loop iterations in parallel. A fixed-point version of the PSR algorithm is also proposed, with a modification to ensure that a correct number of particles is generated. Experiments show that the fixed-point implementation of the PSR algorithm can use as few as 22 bits for representing the weights, when processing 512 particles, while achieving results equivalent to a floating-point SR implementation. Four customized instructions were designed to accelerate the proposed PSR algorithm in Application-Specific Instruction-set Processors. These four custom instructions, when configured to support four weight inputs in parallel, lead to a 73.7 $$\times $$ × speedup over a floating-point SR implementation on a general-purpose processor at a cost of 47.3 K additional gates.

Diffusion in periodic potentials with path integral hyperdynamics

We consider the diffusion of brownian particles in one-dimensional periodic potentials as a test bench for the recently proposed stochastic path integral hyperdynamics (PIHD) scheme [Chen and Horing, J. Chem. Phys. 126, 224103 (2007)]. First, we consider the case where PIHD is used to enhance the transition rate of activated rare events. To this end, we study the diffusion of a single brownian particle moving in a spatially periodic potential in the high-friction limit at low temperature. We demonstrate that the boost factor as compared to straight molecular dynamics (MD) has nontrivial behavior as a function of the bias force. Instead of growing monotonically with the bias, the boost attains an optimal maximum value due to increased error in the finite path sampling induced by the bias. We also observe that the PIHD method can be sensitive to the choice of numerical integration algorithm. As the second case, we consider parallel resampling of multiple bias force values in the case of a brownian particle in a periodic potential subject to an external ac driving force. We confirm that there is no stochastic resonance in this system. However, while the PIHD method allows one to obtain data for multiple values of the ac bias, the boost with respect to MD remains modest due to the simplicity of the equation of motion in this case.

Design and Implementation of Flexible Resampling Mechanism for High-Speed Parallel Particle Filters

There are many applications in which particle filters outperform traditional signal processing algorithms. Some of these applications include tracking, joint detection and estimation in wireless communication, and computer vision. However, particle filters are not used in practice for these applications mainly because they cannot satisfy real-time requirements. This paper presents an efficient resampling architecture for parallel particle filtering. The proposed architecture is flexible such that it supports various modes of parallel resampling operations with up to four processing elements. The resampling algorithm is developed in order to compensate for possible error caused by finite precision quantization in the resampling step. Communication between the processing elements after resampling is identified as an implementation bottleneck, and therefore, concurrent buffering is incorporated in order to speed up communication of particles among processing elements. The flexible resampling mechanism is implemented in 0.35 μ m CMOS process and its complexity and performance are analyzed.

High-throughput scalable parallel resampling mechanism for effective redistribution of particles

A novel resampling mechanism for parallel processing of fixed-point particle filtering is discussed. The proposed mechanism utilizes a particle-tagging scheme during quantization to compensate possible loss of replicated particles due to the finite precision effect. Particle tagging divides replicated particles into two groups for systematic redistribution of particles to eliminate particle localization in parallel processing. The mechanism utilizes an efficient interconnect topology for guaranteeing complete redistribution of particles even in case of potential weight unbalance among processing elements. The proposed architecture supports high throughput and ensures that the overall parallel particle filtering execution time scales with the number of processing elements employed.