A parallel color-based particle filter for object tracking

Abstract

Porting well known computer vision algorithms to low power, high performance computing devices such as SIMD linear processor arrays can be a challenging task. One especially useful such algorithm is the color-based particle lter, which has been applied successfully by many research groups to the problem of tracking nonrigid objects. In this paper, we propose an implementation of the color-based particle lter suitable for SIMD processors. The main focus of our work is on the parallel computation of the particle weights. This step is the major bottleneck of standard implementations of the color-based particle lter since it requires the knowledge of the histograms of the regions surrounding each hypothesized target position. We expect this approach to perform faster in an SIMD processor than an implementation in a standard desktop computer even running at much lower clock speeds.

Algorithm and architecture for object tracking using particle filter

“…In our software implementation, we can only achieve 8 frames per second (fps) in a computer with Intel Core Duo 2.66GHz and 2GB RAM. Although many previous works focus on the parallelization of the resampling part, our profiling result is consistent with [1] and [6], showing that histogram building and comparison require more than 90% of the operation time. Therefore in our architecture design, we focus on the acceleration of the histogram building and calculation.…”

“…Recently, [6] proposed to implement the color-based particle filter in a SIMD processor and utilized its line memory to implement the color histogram accumulation in parallel. In spite of its effectiveness in utilizing the hardware resource of a SIMD processor, the capability to compute color histograms in parallel is bounded by the hardware resource, for example the line memory of the architecture.…”

A parallel histogram-based particle filter for object tracking on SIMD-based smart cameras

“…The particles x i correspond to the hypothesized positions of the target, and the measurements z i are given by the histograms of the regions surrounding each particle. The particle likelihoods are computed based on the similarity between the measurements and the target reference histogram [3,5,18]. In order to process the particle regions in parallel, it is necessary to allocate a certain number of PEs to handle each region.…”

“…This paper is an extension of our previous work [17,18] in which we proposed a method for the parallel computation of the particle weights in the color-based particle filter on an SIMD processor. In this paper, we show that our methodology is valid for any histogram-based feature set-we show in detail how the parallel particle filter can employ not only simple color histograms but also complex feature descriptors, specifically the histograms of oriented gradients (HOG).…”

Resource-aware distributed particle filtering for cluster-based object tracking in wireless camera networks

“…The transition noise, u t , is set as a Gaussian noise model. Also, we consider y t as the target object (or reference) feature at time t and h t (x t ) as the extracted feature at a support point x t in the image at time t. In our implementation, we use color histogram features as described in [16,18,59]. We compute the object histogram at the object position on the image coordinate, which is transformed from the object position on the world coordinate frame by the precomputed homography obtained from the camera calibration information.…”