System flowchart

Our result

CAMShift wrong tracking result

Confusion matrix of recognizing grade rate
Kinect-Based Autonomous Training Assistant for Basketball Dribbling

  For beginners, basketball dribbling training requires instructions from coaches. To ensure the training is effective, coaches usually teach details and emphases of each move and correct mistakes right away as a bystander. Then, they might just need training videos to practice autonomously, but the correctness of moves is not able to confirm objectively. Therefore, we present a Kinect-based assistance system for user to train autonomously at home or community centers no matter when. Moreover, our system is able to provide visual feedback of each move in real-time to ensure the correctness.
  We can get user’s skeleton information by using Kinect and Kinect skeletal tracking algorithm. We can also automatically detect basketball and get its location in real-time by improving existing object detect and tracking algorithm with color, depth, and body index images obtained from Kinect. For posture recognition, we separated dribble moves into details and emphases, defined these as independent postures, and built posture database by using a machine learning based tool-Microsoft Visual Gesture Builder. By combining skeleton, ball locations, and posture recognition, we can achieve our goal.
  Our system solves problems such as no coach as bystander or no objective confirmation of correctness of moves. Note that our system does not aim to replace conventional role of coaches, but to allow players and coaches to have more time flexibility. In addition, our system can be applied to serve a basketball team. Coach decides training menu, emphases of moves, and passing standard for each player and gets reports after players finish their training. Therefore, coach can teach more players at once, and player can choose when he/she wants to train in a time limit.

SUMMARY (中文總結):
  使用Kinect以及其骨架追蹤技術可以得到使用者的骨架資訊。而改良現有物件偵測與追蹤方法後我們可以由Kinect取得的影像中即時且自動地偵測與追蹤籃球。另外在姿勢辨識部分,利用Microsoft提供的工具Visual Gesture Builder(簡稱VGB)能將分解後動作細項自定義為各個獨立的姿勢並以機器學習的方法建立姿勢資料庫。我們結合骨架資訊、球的位置以及姿勢辨識來完成自主訓練輔助系統。


  • 計算運球次數

  • 結果展示