Playfield detection in soccer images using prior information
Piya Kaewbuadee, Phatthanaphong Chomphuwiset · 2016
The detection of playfield in sport videos (particularly soccer games) is considered as an initial step for higher semantic analysis such as player tracking and identifying. This work proposes a technique for soccer playfield detection in images. The proposed technique comprises with 4 main steps, i.e. (i) pre-processing, (ii) feature extraction, (iii) classification and (iv) refinement. Color transformation is first carried out, in the pre-processing, to reduce color variation in the images, before feature extraction is performed. This work applies two main types of features, which are color intensity and the texture around pixels. In classification, a generative model is constructed using Naive Bayes algorithm. The model uses a prior information of the regions of playfield in images. Each training image is divided into three horizontal regions to designate the most probable region of the playfield located in the images. After carrying out the classification, a refinement process is implemented. This refinement aims at improving the classification results by utilizing some contextual information of spatial relationships of the pixels in the images using Markov Random Field technique. In addition, medial filter is implemented in this refinement step. Experiments were conducted with 120 images. The results from the experiments showed that the proposed technique provided promising results using intensity features. The prior information increased the performance of the classification. In addition, the refinement technique provided a better result than performing classification alone (without refinement process).