Research On Visual Targets Tracking Algorithms Based On Particle Filter

Behavior detection and tracking of targets in coal mine dangerous regions,which is dependent on Computer Vision,Pattern Recognition and Artificial Intelligence, etc, has become significant application of Intelligent Visual Surveillance(IVS) in coal mines. In mine IVS systems, targets tracking is a key technology and directly determines its performance. Tracking not only provides accurate positions of targets and data sources for further analysis, but also optimizes the targets detection.Particle filter, using a great deal of stochastic weighted samples to approximating Bayesian filter,has the great advantage of not being subject to the assumption of linearity or Gaussianity of the model in targets states estimation. To improve the robustness of particle filter based tracking algorithms in coal mines, the research is focused on the four major aspects of particle filter: observation models and multi-cues integration strategy, proposal distribution, re-sampling algorithms and Multi-targets Tracking based on particle filter.The main research work includes:A novel unscented particle filter algorithm (ISUKF-PF) has been proposed, using iterated minimal skew simplex UKF (ISUKF) as the proposal distribution. In this paper, statistical liner error propagations were obtained by ISUKF;And we have derived the IEKF iterated equations by replacing the system model Jacobian matrix with statistical liner error propagation terms; Then the states mean and covariance have been iterated and updated by the IEKF iterated equations to be convergent to the state MAP estimation for near zero-residual. The results show that the ISUKF achieved the more overlap regions of prediction samples and peak zones of observation likelihood and increased the accuracy of state estimating in nonlinear system.A novel unscented particle filter(UPF) algorithm has been proposed for object-tracking in coal mines, using ISUKF as proposal distribution and an adaptive multi-cues fusion modal as the observation modal. ISUKF generate prediction samples adjusted to high likelihood area in state-space. Then, Optical flow histogram was proposed and observation model based on multi-cues fusion was implemented by integrating optical flow with color. Adaptive strategy of observation modal weights was implemented by adjusting the contribution rate of single-cue observation modal. Consequently, reliability of an observation modal adjusted to changes of object characteristics accordingly. Finally, a function of sample compensation was proposed to handle particle diffusion due to failure of observation modal. The results show that the tracking algorithm is an effective solution to tracking failure due to invalidation of observation modal in coal mines(complex background).A new method, named layered transacting MCMC-Resampling algorithm, was proposed. When the effective sample size is below a fixed threshold, particles are dived into two sample subsets according to their individual weights. Mutation operator and PSO, which considered as transition kernels of MCMC, applied to sample subsets respectively. Then an acceptance-rejection rule of Metropolis-Hastings algorithm is applied to generate the Markov Chain with the stationary distribution which is equivalent to target posterior density. The results show that the proposed method is superior to the other resampling algorithms both in accuracy and convergence speed.A new object matching algorithm based on object-regions in multiple cameras was proposed. The scale invariant features of object-region were extracted, which used difference-of-gaussian (DOG) algorithm. The region was represented by PCA-SIFT descriptor. Through comparing the matching values of object-regions, objects matching was completed in object-regions matching algorithm; And then, wrong regions were deleted by the fundamental matrix. The results show that the matching algorithm can deal with object matching in coal mine(low illumination level,complex target type) based multiple cameras.An improved JPDA algorithm,approximating confirmation matrix of JPDA by measurements Bhattacharya coefficients,was proposed . Track birth and death was obtained by comparing Bhattacharya coefficients of measurements in a tracking gate. Then Bhattacharya means matrix of measurements area was created and elements of confirmation matrix were optimized to eliminating association events with low association probability. The results show that the proposed method reduced the computational complexity as well as increased precision of estimation.