Abstract:Recently, deep learning methods have been applied in many real scenarios with the development of convolutional neural networks (CNNs). In this paper, we introduce a camera-based basketball scoring detection (BSD) method with CNN based object detection and frame difference-based motion detection. In the proposed BSD method, the videos of the basketball court are taken as inputs. Afterwards, the real-time object detection, i.e., you only look once (YOLO) model, is implemented to locate the position of the basketball hoop. Then, the motion detection based on frame difference is utilized to detect whether there is any object motion in the area of the hoop to determine the basketball scoring condition. The proposed BSD method runs in real-time with satisfactory basketball scoring detection accuracy. Our experiments on the collected real scenario basketball court videos show the accuracy of the proposed BSD method. Furthermore, several intelligent basketball analysis systems based on the proposed method have been installed at multiple basketball courts in Beijing, and they provide good performance.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to automatically detect the scoring situation in basketball games. Specifically, the author proposes a camera - based basketball score detection (BSD) method, which combines the object detection technology of convolutional neural network (CNN) and the motion detection technology of frame difference method. Through this method, the scoring situation can be detected in real - time from basketball game videos, thereby reducing the need for human scorers and improving the automation level of game records.
### Background and Motivation of the Paper
With the development of artificial intelligence, computer vision and data - mining technologies, many intelligent methods have been applied in the sports field. On the one hand, some methods are used to predict the results of games, helping people analyze sports events from a macro level; on the other hand, some methods are used to analyze specific behaviors on the field, helping people understand sports events from a micro level. However, in amateur basketball games, it is not practical to arrange scorers or score sheets, so automated score detection and counting are of great significance for basketball court applications.
### Proposed Method
The paper proposes a camera - based basketball score detection method, which mainly consists of two parts:
1. **YOLO Basket Detector**: Use the YOLO model to locate the position of the basket on the basketball court.
2. **Frame Difference Method Score Detector**: Use the frame difference method to detect whether there is object movement in the basket area, thereby judging whether a score is made.
### Method Details
- **YOLO Basket Detector**:
- Take the first frame of the input video as the reference frame to determine the position of the basket.
- The loss function includes classification loss, center coordinate loss, width - height coordinate loss and confidence loss.
- Classify through the binary cross - entropy loss function, the center coordinate loss and the width - height coordinate loss are used for precise positioning of the basket, and the confidence loss is used to evaluate the accuracy of the detection box.
- **Frame Difference Method Score Detector**:
- After locating the position of the basket, perform frame difference method processing on subsequent frames to detect whether there are moving objects in the basket area.
- Convert the frame difference result into a binary image by setting a threshold, extract the largest moving area after connected - component analysis, and judge whether a score is made.
### Experimental Results
- **Data Set**: A data set containing 4,000 pictures was collected, of which 3,500 were used to train the YOLO basket detector and 500 were used to test.
- **Experimental Platform**: Based on Nvidia GTX TitanX GPU and Intel i7 - 9750 CPU.
- **Performance Evaluation**:
- The YOLO basket detector can reach 30 FPS on the Nvidia GTX TitanX GPU.
- The average precision (AP50) of basket detection is 92.59%.
- The entire BSD method was tested on five long videos, with a total of 44 scoring moments, and the accuracy rate of score detection was 88.64%.
### Conclusion
The paper proposes a camera - based basketball score detection method, which can achieve a satisfactory score detection accuracy while processing basketball game videos in real - time. Future research will further expand the scale of the data set, apply more efficient and effective detection models, and add more basketball analysis functions.