Multi-Object Tracking for Indoor Sports Broadcasting

Abstract

Tracking objects on camera acquired images is one of the important issues in computer vision. By analyzing and tracking objects, it can be used in various fields such as security surveillance, motion analysis of tracked objects. In this paper, we propose a real-time object tracking method for indoor sports broadcasting service. The image sequence is acquired with a top-view camera installed on the ceiling of the indoor stadium. Using the acquired images, the players are classified and tracked to calculate the position information, and the PTZ camera captures the image of the corresponding player using the calculated position information. For the indoor broadcasting service, the position of the players should be tracked in real time, and the players should be able to track even the rapid movement and overlap. In addition, the location of the tracking object should be tracked rather than missed. To solve these problems, we propose GPGPU based real - time tracking algorithm. The background image is extracted through the background modeling based on the image acquired from the fixed Top-view(global) camera. The object region is identified using the difference between the extracted background image and the inputted input image. To analyze the characteristics of objects, we analyze the characteristics of objects using color information and shape information. The color information of the athletes is analyzed by Hue-Saturation Histogram(HS Histogram) using team uniform colors. The shape information of the players is analyzed using the Histogram of Oriented Gradients (HOG), which best analyzes human form information. These two features are used to select the reference features of the object to be tracked. The position of the object in the input image is estimated by the Particle Swarm Optimization (PSO) based on particle filter which is robust against irregular motion, and the position of the tracking object is calculated by calculating the region most similar to the characteristic of the reference object. The PSO, which predicts the position of the object using the generated particles in the search area, calculates the optimal solution by repeatedly calculating the similarity with the reference feature. In this case, since the HS histogram is generated by the number of particles and the number of repetitions, the calculation speed is increased. In order to solve this problem, we improved the calculation speed by using Integral Histogram similar to Integral Image. The integral histogram can be calculated quickly by calculating the histogram from the origin to the position of each pixel in advance by calculating the histogram of the specific region by using the histogram of the four positions. After tracking the position of the object using the PSO, if tracing fails such as occlusion, draft, etc., the final position is corrected by analyzing the trajectory of the previous tracking positions. For the real time service, the feature analysis, the integral histogram, and the algorithm of the PSO with high computational complexity were accelerated using the CUDA based GPGPU. Experimental results show that the processing speed is 8 fps ~ 25 fps according to the size of the integral image and there is no problem in real time processing.