Norikazu Ikoma

Extended object tracking using particle techniques

In This work we consider the problem of extended object tracking. An extended object is modelled as a set of point features in a target reference frame. The dynamics of the extended object is formulated in terms of the translation and rotation of the target reference frame relative to a fixed reference frame. This leads to realistic, yet simple, models for the object motion. We assume that the measurements of the point features are unlabelled, and contaminated with a high level of clutter, leading to measurement association uncertainty. Marginalising over all the association hypotheses may be computationally prohibitive for realistic numbers of point features and clutter measurements. We present an alternative approach within the context of particle filtering, where we augment the state with the unknown association hypothesis, and sample candidate values from an efficiently designed proposal distribution. This proposal elegantly captures the notion of a soft gating function. We demonstrate the performance of the algorithm on a challenging synthetic tracking problem, where the ground truth is known.

Maneuvering target tracking by using particle filter

The aim of this research is to track a maneuvering target, e.g. a ship, an aircraft, and so on. We use a state-space representation to model this situation. The dynamics of the target is represented by a system model, firstly in continuous time, though a discretized system model is actually to be used in practice. The position of the target is measured by radar, and this process is described by a nonlinear observation model in polar coordinates. To follow abrupt changes in the targets motion due to sudden operations of the acceleration pedal, braking and steering, we propose the use of heavy-tailed non-Gaussian distribution for the system noise. Consequently, the model we use is a nonlinear non-Gaussian state-space model. A particle filter is used to estimate the target state of the nonlinear non-Gaussian model. The usefulness of the method is shown by simulation.

Anti-occlusion observation model and automatic recovery for multi-view ball tracking in sports analysis

The 3D position of the ball plays a crucial role in professional sport analysis. In ball sports, tracking of balls precise position accurately is highly required, whose performance is affected by inaccurate 3D coordinates and occlusion problem. In this paper, we propose anti-occlusion observation model and automatic recovery by 3D ball detection based on multiview videos to track the ball in 3D space. The anti-occlusion observation model evaluates each cameras image and eliminates the influence of the cameras in which the ball is occluded. The automatic recovery method detects the balls 3D position by homography relation of the multi-video and generates a new distribution to initiate the tracker when tracking failure is detected. Experimental results based on the HDTV video sequences, which were captured by four cameras located at the corners of the court, show that the success rate of the 3D ball tracking achieves 99.14%.

GPGPU implementation of visual tracking by particle filter with pixel ratio likelihood

Visual tracking by particle filter with pixel ratio in a region of interest for likelihood computation has wide range of applications despite of its simple algorithm. A GPGPU (General Purpose computation on Graphics Processing Unit) implementation of the visual tracking in parallel computation has been proposed in this paper. Algorithm of the tracker has almost fully been implemented in CUDA framework. Difference of the proposed algorithm from the full algorithm is a reduction of image size from the original image in order to deal with multiple images for likelihood computation in limited size of constant memory of the GPU hardware. Performance of the proposed method achieves 30 fps (frame per second) for specific colored object tracing task and more than ten frame per second for a task of hands tracking of a car driver operating a steering.

Tracking of maneuvering target by using switching structure and heavy-tailed distribution with particle filter method

The tracking problem of maneuvering target with an assumption that the maneuver is unknown and its acceleration has some abrupt changes is treated by formulating a general (nonlinear, non-Gaussian) state space model with the system model to describe the target dynamics and observation model to represent a measurement process of the target position. The Bayesian switching structure model, which includes a set of possible models and switches among them, is used to cope with the unknown maneuver. The heavy-tailed uni-modal distribution, e.g. Cauchy distribution, is also used for the system noise to accomplish good performance of tracking both the constant period and abrupt changing time point of acceleration. The Monte Carlo filter, which is a kind of particle filter that approximates state distribution by many particles in state spare, is used for state estimation of the model. A simple simulation study shows an improvement of performance by the proposed model comparing with a Gaussian case of the Bayesian switching structure model.

Adaptive order selection with aid of genetic algorithm

A method to estimate a nonstationary power spectrum with adaptive selection of autoregressive order is proposed. Time-varying PARCOR (partial autocorrelation coefficient) and AR (autoregressive) order are estimated from time series data. The data are assumed to be observations of vibration that contain abrupt change of spectrum due to arrivals of different signal, structural changes of vibrating object, etc. The model that consists of an autoregressive model with time-varying PARCORs and time-varying order is used. The time-varying PARCORs are estimated by a Monte Carlo filter, and the time-varying order is estimated by genetic algorithm. An application to analysis of seismic wave data is reported.

Hands and Arms Motion Estimation of a Car Driver with Depth Image Sensor by Using Particle Filter

Estimation of hands and arms motion in a car driving by utilizing a depth image sensor, specifically, KINECT of Microsoft Xbox 360, has been proposed. Compared with conventional researches using ordinary vision sensor, depth sensor provides rich information for the hands and arms in the scene. Especially, arms’ regions detected by the depth sensor have been utilized to estimate the hands and arms motion more accurately than the conventional researches. As well as the increasing accuracy of the hands and arms region extraction, this paper proposes to incorporate some particles intentionally switching the left and the right of the hands in a framework of particle filter. This idea reduce the mistaken (opposite) determination of left and right and it will increase the opportunity to recover automatically from the opposite determination. Experiments over the recorded videos of vision and depth under a driving simulator environment show the efficiency of the proposed method.

Multiple pedestrians tracking with composite sensor of laser range finder and omni-directional camera by SMC implementation of PHD filter

Multiple pedestrians tracking with composite sensor of laser range finder and omni-directional camera having wide field of view has been proposed based on Sequential Monte Carlo (SMC) implementation of Probability Hypothesis Density (PHD) filter in Track Before Detect (TBD) approach. The composite sensor has been fixed on the flat ground with its laser measurement plane crossing body of target pedestrians. Tracking has been done for multiple pedestrians walk around the sensor with cluttered static scene. Real-time tracking system has been developed and experimental results demonstrate the performance of the proposed method.

Likelihood adjustment among multiple targets for particle dependent tracking in particle filters

A problem arising at multiple target tracking with particle filters typically in vision has been claimed and a likelihood adjustment method has been proposed. First, classify tracking methods by particle filters into two categories, detection first tracking and particle dependent tracking. Then this research focus on the particle dependent tracking. It involves the problem in case of multiple target tracking that difference of likelihood among target leads to unintended convergence of particles to one target. This is a phenomenon in particle filters that particles prefer easier target having large likelihood value than the difficult target to track having small likelihood value. To overcome this problem, the author proposes to adjust the likelihood among the targets by taking difference in log-likelihood to local maximum of them in each target. Performance of the proposed method has been shown in visual target tracking experiment based on color region.

Sound target tracking in 3D using particle filter with 4 microphones

We propose a method to detect and to track a moving sound target in 3D using particle filter with four microphones, which are allocated not being on the same plane. We develop an elaborated state space model where state represents 3D location of the sound target, and observation is 6 sound directions obtained by 6 pairs of the 4 microphones. Each observed sound direction is calculated from TDOA measurements of corresponding microphone pair by taking maximum of cross correlation function of two sound signals. Through a simulation experiment of single sound target tracking in noisy environment, we show performance of proposed method.