基于YOLO算法实现足球运动实时分析(附源码)

大家好，我是小F～

今天给大家介绍一个计算机视觉实战的项目。

这个项目使用YOLO（目前最好的AI物体检测模型之一）来检测跟踪视频中的球员、裁判和足球。

使用Kmeans进行像素分割和聚类，根据球员的T恤颜色将他们分配到各个球队。

有了这些信息，就可以衡量一支球队在比赛中的控球率。

此外使用光流测量帧之间的摄像机移动，使得能够准确测量球员的移动。

通过透视变换来表示场景的深度和透视，使我们能够以米而不是像素为单位测量球员的移动。

最后，我们将计算球员的速度和覆盖的距离。

需要安装Python3，推荐使用Python3.10。

下面是一些Python库的依赖。

ultralytics
supervision
OpenCV
NumPy
Matplotlib
Pandas

主程序代码如下~

from utils import read_video, save_video
from trackers import Tracker
import cv2
import numpy as np
from team_assigner import TeamAssigner
from player_ball_assigner import PlayerBallAssigner
from camera_movement_estimator import CameraMovementEstimator
from view_transformer import ViewTransformer
from speed_and_distance_estimator import SpeedAndDistance_Estimator


def main():
    # Read Video
    video_frames = read_video('input_videos/08fd33_4.mp4')

    # Initialize Tracker
    tracker = Tracker('models/best.pt')

    tracks = tracker.get_object_tracks(video_frames,
                                       read_from_stub=True,
                                       stub_path='stubs/track_stubs.pkl')
    # Get object positions 
    tracker.add_position_to_tracks(tracks)

    # camera movement estimator
    camera_movement_estimator = CameraMovementEstimator(video_frames[0])
    camera_movement_per_frame = camera_movement_estimator.get_camera_movement(video_frames,
                                                                              read_from_stub=True,
                                                                              stub_path='stubs/camera_movement_stub.pkl')
    camera_movement_estimator.add_adjust_positions_to_tracks(tracks, camera_movement_per_frame)

    # View Trasnformer
    view_transformer = ViewTransformer()
    view_transformer.add_transformed_position_to_tracks(tracks)

    # Interpolate Ball Positions
    tracks["ball"] = tracker.interpolate_ball_positions(tracks["ball"])

    # Speed and distance estimator
    speed_and_distance_estimator = SpeedAndDistance_Estimator()
    speed_and_distance_estimator.add_speed_and_distance_to_tracks(tracks)

    # Assign Player Teams
    team_assigner = TeamAssigner()
    team_assigner.assign_team_color(video_frames[0],
                                    tracks['players'][0])

    for frame_num, player_track in enumerate(tracks['players']):
        for player_id, track in player_track.items():
            team = team_assigner.get_player_team(video_frames[frame_num],
                                                 track['bbox'],
                                                 player_id)
            tracks['players'][frame_num][player_id]['team'] = team
            tracks['players'][frame_num][player_id]['team_color'] = team_assigner.team_colors[team]

    # Assign Ball Aquisition
    player_assigner = PlayerBallAssigner()
    team_ball_control = []
    for frame_num, player_track in enumerate(tracks['players']):
        ball_bbox = tracks['ball'][frame_num][1]['bbox']
        assigned_player = player_assigner.assign_ball_to_player(player_track, ball_bbox)

        if assigned_player != -1:
            tracks['players'][frame_num][assigned_player]['has_ball'] = True
            team_ball_control.append(tracks['players'][frame_num][assigned_player]['team'])
        else:
            team_ball_control.append(team_ball_control[-1])
    team_ball_control = np.array(team_ball_control)

    # Draw output 
    ## Draw object Tracks
    output_video_frames = tracker.draw_annotations(video_frames, tracks, team_ball_control)

    ## Draw Camera movement
    output_video_frames = camera_movement_estimator.draw_camera_movement(output_video_frames, camera_movement_per_frame)

    ## Draw Speed and Distance
    speed_and_distance_estimator.draw_speed_and_distance(output_video_frames, tracks)

    # Save video
    save_video(output_video_frames, 'output_videos/output_video.avi')


if __name__ == '__main__':
    main()

项目使用了以下模块： 

• YOLO：AI物体检测模型

• Kmeans：像素分割和聚类来检测T恤颜色

• 光流：测量相机运动

• 透视变换：表示场景深度和透视

• 计算每位球员的速度和距离

项目源码，公众号后台回复：「足球分析」，即可获得。

万水千山总是情，点个 👍 行不行。

推荐阅读

···  END  ···