APPLICATIONS Soccer Players Detection Using GDLS Optimization and Spatial Bitwise Operation Filter

The advancement in computer vision technology in order to help coach creates strategy has been affecting the sport industry evolving very fast. Players movement patterns and other important behavioral activities regarding the tactics during playing the game are the most important data obtained in applying computer vision in Sport Industry. The basic technique for extracting those information during the game is player detection. Three fundamental challenges of computer vision in detecting objects are random object’s movement, noise and shadow. Background subtraction is an object’s detection method that used widely for separating moving object as foreground and non moving object as background. This paper proposed a method for removing shadow and unwanted noise by improving traditional background subtraction technique. First, we employed GDLS (Gradient Decent with Line Search) algorithm to optimize background-foreground separation. Then, we did filter shadows and crumbs-like object pixels by applying digital spatial filter which is created from implementation of digital arithmetic algorithm (bitwise operation). Finally, our experimental result demonstrated that our algorithm outperform conventional background subtraction algorithms. The experiments result proposed method has obtained 80.5% of F1-score with average 20 objects were detected out of 24 objects.


I. INTRODUCTION
Automatic multiple objects detection in soccer has played an important role in game monitoring system.
Due to high demand of implementing information technology in metric measurement began enormously emerging, research on player's automatic annotation system is flourishing astonishingly. It helps performance analysts and coaches to plan future strategies. Nowadays, this technique has played a crucial role in premiere leagues due to its impact on the competition. The need of sophisticated system which can analyze both players behavioral pattern accurately and their annotation motion during the game can be a valuable competition strategy. Therefore, current game play of international soccer are much more developed than before when such system was not yet implemented. When in each team having the same framework of it, the output will be having same manner. Thus, it will create fierce competition battle between the teams. Furthermore, team without using this strategy will have small change of superiority. Some previous works have the focus on developing automatic detection and tracking algorithm using background subtraction technique [2], [3], [4]. Background subtraction technique is the process based on field image processing technique wherein a foreground is separated from its background in order to further processing. There are two categories of detecting and tracking algorithm based on [1], they were traditional model and recent approach models.
Traditional model deal with the basic characteristic and relatively prone to implement without complex parameter setting. It began from basic model approaches using running average, median and histogram approaches to build background mathematical model [2] [3] [4]. Even though using basic computation model with a focus on an object as random variable which are solved by conventional probabilistic methods, it requires less sources than any other recent model has. However, based on research [1], it can be concluded that, its detection quality is lower than the recent model has.
Furthermore, recent models handled challenges such as real-time operation [5], illumination [6], noise cancelation [7], shadow casting [8], and the false alarm [9] in a better way. In other studies, statistical model of background modeling are being modeled from Gaussian distribution [10] [11], support vector model [12], and subspace learning [12] [13]. Implementation of this traditional approaches has been set by the following models such as cluster-based [14], neural networks basis model and estimation-based model [15]. Recent approach works by modeling the object based on subspace approach like (Robust Principle Component Analysis) RPCAapproaches [16], decomposition of sparse matrix [17], subspace tracking based approach [1], low rank minimization [18] [19], robust tensor models [20], outlier detections and transform domain model approach such as FFT, DCT, Walsh, Wavelet, and Hadamard transform approaches [1]. This means, it will add more load in computational effort than any other unification methods.  (2), image difference (3) between background maintenance process (1) will send to find optimum threshold (4), which it employs GDLS algorithm. In Bitwise operation we did concatenate three layers HSV to get filtered binary frame then after output was resulted crumps residue was wiped out by morphological operation. Fig. 2. Background subtraction process. Video Frames as input is responsible to fetch N frames for maintain the background. Foreground detection is generated by initializing background maintenance process By observing that two categories, our proposed method is combining above issues, which is using traditional model for less complexity and combining using optimization algorithm and shadow casting technique to achieved higher result like recent model has. We found that binary filter which running in array platform can be used to simplify computational load. We also hypothesize that some other challenges in player's shadow from illumination distribution and noise artefacts from codex or background subtraction error can be solved by spatial bitwise operation filter with an improvement from GDLS algorithm. GDLS or Gradient Decent by Line Search is an algorithm to find the maximum difference between the background and the foreground.

II. PROPOSED METHOD
The aim of our modified background subtraction method is to find best threshold for binarization and to filter shadow and grumps-like artifact which affect object's brightness level. We grouped our works by ADHI D WIBAWA ET AL. / J. DATA SCI. APPL. 2019, 2 (1): 1-10 Soccer Players Detection Using GDLS Optimization and Spatial Bitwise Operation Filter 3 two categories, the first is background subtraction and the second is spatial filter using bitwise operation for filtering unwanted noise and shadows. Fig 1 shows the flow work of our proposed method

A. Background Subtraction
In technical definition, background subtraction technique is taking the difference between a background image and object image. Therefore object is separated with its background. In order to separate the background, we need to understand the spatial features on data pixel characteristics [3]. Five features that are commonly used for segmenting or clustering the object are: color, edge, stereo, motion and texture features [1]. Based on our observation in Table 1 the change of illumination (lighting), waking object, sleeping object, and shadows become apparent that those items did not radically change the object background, therefore we hypothesize that using color-based feature for doing object segmentation is fair. However, in soccer which both teams are distinguished by their jersey color, so color-based segmentation technique is used for separating foreground-background.
We found that HSV color has more information about brightness level than RGB color [3]. Therefore, we model an image by ( , ) where and are the pixel coordinate locations and has three dimension which as , , and regarding with three dimensional HSV color-space. The brightness (. ) is used to level of an object based on point operation. It is given by ′( , ) ← ( ( , )). This function affects natural pixel of background and foreground . Ideal background subtraction is aiming the best finding of background model which suit with actual image or we can say with: From [1], we emphasize that: prior designing suitable methods, we should list key challenge of datasheet features provided that we could best address the key challenges. Table 1 below will help to conduct this activity. Based on theoretical review [1], we found that traditional spatial based operation algorithm is enough, since its datasheet has no dynamic background-foreground changing. Consequently, we choose improved-traditional-background subtraction-based technique as our building block of process method.
Image difference of modeled background and image testing is given by: In (1), image difference ( , ) pixels are being thresholdedby Γ = [ 1 2 3 ]. In order to get binary masking (Γ), we define as detected objects of-. Obtaining higher accuracy of leads to higher Fmeasurement. Choosing ( , ) is very crucial, because bad choice of the threshold will lead to false positive detection. Detailed process of background subtraction can be seen at Fig 2. There are four process which are background maintenance, background initialization, foreground detection and masking process.

1) Background Initialization
There is a challenge in traditional background subtracting method where more than half of the required training images are foreground. As a result, the positive background can only be generated partially (half of which is foreground). Therefore, the number of initialization images need to be multiplied until we get full background information. Fig 2 showed a causal relationship of background subtraction process. Background maintenance will maintain the change of background initialization, while background initialization will adapt to waking or moving background by taking an average of sampled frames. As a result this will produce an image that contains only the background without foreground

2) Foreground Detection
In this process, we have done foreground-background separation, by subtracting output background maintenance and input frame. The result is an output image that has close-to-zero value of background pixel and -distance value of object pixel. The aim of this process is extracting foreground from image input with background model ∅. Since we need to maintain ( , ) to be optimum, background maintenance formula is: Learning rate is constant in (0,1). Since the background is the static and luminance distribution is constant in each frame. Thus, = 0 so that (2) became: In addition, we compute background as model approach * for the entire role of image input. Based on our research trial gathering N>50 frames for calculating is enough, because stationary cameras with constant illumination were being placed in the field.

3) Foreground Masking
The proposed method is using optimization method Gradient Decent with Line Search algorithm (GDLS). Our problem statement is to find arg Γ (Γ). Possibly subject to constant Γ. Where finding min (Γ) is equal to find best formation of Γ = [ 1 2 3 ] and Γ ∈ ℝ . (Γ) → ‖ (Γ) − (Γ)‖ with (Γ) = (Γ) + (Γ). We can say that (Γ) is the residue from equation (Γ). In this experiment, we define image as ground truth From calculus, we recall that minimum of (Γ) must lie at point when (Γ) (Γ) = 0 and it will automatically maximize ( | ). We iterate and sort it based on the number of dimension. Each optimal value is found using GDLS. For each dimension step size for each t time, will produce variable . This variable will be updated for future comparison in (Γ) < (Γ ). Note that whenever the weight of previous − ( −1) ≈ ‖ (Γ) − (Γ)‖ are specified, we have a complete rule of future comparison. This means that the difference value between ground truth and foregorund is clearly separated.

B. Spatial Filter using Bitwise Operation
Bitwise operator is found based on spatial information form (Γ). Since this is three color represented in HSV color format, then finding best threshold should be observed by these color. Saturation is capturing distance between origin and Hue value. On the other hand, saturation is the radius between origin and hue point. Hue is the angle around the central vertical axis. Value is the height of cylindricalcoordinate from the origin [21].
Output H, S and V were derived from by applying threshold Γto (Γ). The spatial filter does not need progressive computation because it was applying the array-to-array computation. Progressiveness in computation will lead to more memory allocation, consequently, it result in higher complexity than direct computation.

C. Dataset
This proposed method was compared with four widely used technique PCA based and spatial smoothing model (PCASM) encourage by [9], Gaussian Mixture Model (GMM) [11] [8], Dynamic thresholding for shadow removal Algorithm (DTSR) [22] and Point Operation Background Subtraction (POBS). All object detector are picked based on [1] recommendation in overcoming key challenges.
Typical datasets are consist of 26,222 frames of 40 minutes video h264 format provided by [23]. Each frame wide 2000 by 4500 pixels. The datasheet has some characteristics which can be our challenge, are summarized bellow:  Skewed lighting distribution, sees Fig. 5 (left).  Dynamic random foreground movement.  Shadow is detected as foreground.  Since the video is constructed by the panoramic synthetic video camera, then object's shape is changing in curved area.  Occlusion is occurred and  The object will appear intermittent because it was located in distance. ADHI D WIBAWA ET AL. / J. DATA SCI. APPL. 2019, 2 (1): 1-10 Soccer Players Detection Using GDLS Optimization and Spatial Bitwise Operation Filter 7

D. Evaluation Matrix
Our experiments were using standard metric for the testing performance of object detection algorithm [24]. The quantitative comparison is performed using F1-score measurement which developed by True Positive (TP), False Negative (FN), and False Positive (FP). Where as an input image and Ω as ground truth image.

III. RESULTS AND DISCUSSION
We explained about technical flow of building proposed method in chapter II previously. In this session we will discuss the result by implementing given methods. Fig 1 (2) shows the image difference result. Based on the experiment, we acquire that all channels have their own characteristics. In image difference , it preserves grayscale value greater than 20% of lower parts, therefore we can see the field line and bright jersey have more dominant than the field. While, preserves greater field line value and some parts of dark jersey players. We also got preserves higher amount of total players value and their shadows. Thus, we could combine threshold for[ , , ] to produce a better form of foreground detector. However, some testing regarding the algorithm and the result need to be done to measure the effectieveness in applying new threshold method. Critical threshold value is also need to be determined by further experiment.
We proposed GDLS optimization technique for finding the threshold. We performed training process in all the training sequences, and as a result, we achieved threshold Γ with minimum . Based on our research using Algorithm 1 approaches, we got Γ = [0.1 0.5 0.1]. In each 30 frames processing, we found this threshold after total 6,480 times of iteration. The bitwise operation will filter thresholded output. All described stages of each logic gate's result can be found in  Fig 4(f) shows that the shadows can be filtered effectively by given filter.
However, there were numbers of crumbs-like artifacts that sometimes occur (Fig 3 (7)) which considered as noise. Even tough there were still many noise artifacts which lead to false positive detection, it did not affect the whole detections. F1-score in Table II showed that our result was 0.805 superior to other existing algorithms. TP, FN, and FP were 0.731, 0.1754 and 0.1793 respectively. For benchmarking our result, we compared against GMM, DTSR, PCASM, and POBS and they were resulting F1 score 0.281, 0.456, 0.648 and 0.754 respectively. ADHI D WIBAWA ET AL. / J. DATA SCI. APPL. 2019, 2 (1): 1-10 Soccer Players Detection Using GDLS Optimization and Spatial Bitwise Operation Filter 8 Fig. 6. Illustration of ROC curve, proposed method lies inside diagonal isocost line, which representing good equal cost-performance [25] Skewed lighting distribution of field will result gradual change of brightness from bright to dark and it was affecting object's pixel value (Fig 7-left) to become more likely close to pixel's field. Those objects were hard to distinguish and it will lead detector produce miss detection. Fig 7-middle showed our error detection coded object as background. Fig 7-right showed recovery process of suffered object processed by morphological operation. Another similar illumination challenge was shown in Fig 7 (left) and our proposed method were successfully subtracted background into foreground. In other word, traditional background subtraction by using equation-(4) was proven successfully for separating background from dynamic foreground. Dataset were using camera's panoramic format [23], therefore objects will change in shape, when they were locating on curved frame. This phenomenon can be seen in Fig 3. This condition will affect image representation, so that object location will be differ than it should be. Consequently for accurately determine the object, coordinate points required coefficients of camera curvature constraint. Fig.7. Left: image source, middle: skewed distribution of light causing detector unable to detect some body parts, right: object recovery using morphology operation Grumps-like artifacts noise appear because there were error in video coding format or some moving pixels which coded as foreground. It can be seen at Fig 7 (middle). By employing median spatial filter, these noises can be eradicated.
Proposed method was locating in AUC (Area Under Curve) (Fig 6) which is an area under the trapezoid curve between the blue and green line. From that curve, proposed method has been classified as good acceptable method [25]. Therefore, we can conclude that in Table II proposed method is superior to other methods is justifiable. In Fig 3 we can see that there were two methods, which stated as the acceptable method they were POBS 75.4% and proposed method, whereas PCASM was in barely acceptable in 64.8%