GPU-based chromatic co-occurrence matrices for tracking moving objects
- PDF / 3,807,579 Bytes
- 14 Pages / 595.276 x 790.866 pts Page_size
- 71 Downloads / 250 Views
ORIGINAL RESEARCH PAPER
GPU‑based chromatic co‑occurrence matrices for tracking moving objects Issam Elafi1 · Mohamed Jedra1 · Noureddine Zahid1 Received: 3 July 2018 / Accepted: 13 April 2019 © Springer-Verlag GmbH Germany, part of Springer Nature 2019
Abstract Generally, a good tracking system requires a huge computation time to localize, with accuracy, the target object. For real-time tracking applications, the running time is a critical factor. In this paper, a GPU implementation of the chromatic co-occurrence matrices (CCM) tracking system is proposed. Indeed, the descriptors based on CCM help to improve the accuracy of the tracking. However, they require a long computation time. To overcome this limitation, a parallel implementation of these matrices based on GPU is incorporated to the tracker. The developed algorithm is then integrated into an embedded system to build a real-time autonomous embedded tracking system. The experimental results show a speed up of 150% in the GPU version of the tracker compared to the CPU version. Keywords Chromatic co-occurrence matrices · Particle filter · Real time · GPU · Embedded system
1 Introduction In the last decades, the tracking field became a very important domain for many applications such as surveillance [1–3], driving assistance [4, 5] and robot control [6, 7]. Before tracking an object, a detection phase is required. Many methods are used to detect a moving object in a given scene. The most used among them is the popular background subtraction method [8]. Once the object is detected, a tracking phase is performed by establishing a frame to frame update of the object location. At this stage, several methods based on learning systems [9–12], Histogram of Oriented Gradient (HOG) [13] or optical flow [14] can be used. Most existing tracking methods use the histogram as a descriptor of the target object. However, this descriptor cannot represent efficiently a target object as it finds some difficulties to distinguish objects with the same color. Among the proposed solutions are the chromatic co-occurrence matrices (CCM) [15–17]. The co-occurrence matrix uses the information of texture and color to represent the target object. Thus, even if the colors of the objects are similar, * Issam Elafi [email protected] 1
Laboratory of Conception and Systems (Electronics, Signals, and Informatics), Faculty of Science, Mohammed V University, Rabat, Morocco
their moments are different. Initially, the co-occurrence matrices are defined for gray-level images called “Gray Level Co-occurrence Matrix” (GLCM) [18] and then successfully adopted for chromatic images. However, CCMs are computationally expensive than histograms, which prevents their usage in several domains and especially for realtime tracking. This paper proposes a real-time autonomous embedded tracking system with a parallel implementation of the chromatic co-occurrence matrices to address the computation time limitation. The Graphics Processing Unit (GPU) is a special set of processors designed to accelerate co
Data Loading...
