[Paper Review] Player Tracking and Identification in Ice Hockey

Kanav Vats, Pascale Walters+3 more|arXiv (Cornell University)|Oct 6, 2021

Video Analysis and Summarization4 citations

TL;DR

This paper presents an automated system for tracking and identifying players in NHL broadcast videos using a three-stage pipeline: multi-object tracking, team classification, and player identification via a temporal 1D CNN. The system achieves 82.8% player identification accuracy by leveraging tracklets, team labels, and NHL roster data, with a MOTA of 94.5% and IDF1 of 62.9%.

ABSTRACT

Tracking and identifying players is a fundamental step in computer vision-based ice hockey analytics. The data generated by tracking is used in many other downstream tasks, such as game event detection and game strategy analysis. Player tracking and identification is a challenging problem since the motion of players in hockey is fast-paced and non-linear when compared to pedestrians. There is also significant camera panning and zooming in hockey broadcast video. Identifying players in ice hockey is challenging since the players of the same team look almost identical, with the jersey number the only discriminating factor between players. In this paper, an automated system to track and identify players in broadcast NHL hockey videos is introduced. The system is composed of three components (1) Player tracking, (2) Team identification and (3) Player identification. Due to the absence of publicly available datasets, the datasets used to train the three components are annotated manually. Player tracking is performed with the help of a state of the art tracking algorithm obtaining a Multi-Object Tracking Accuracy (MOTA) score of 94.5%. For team identification, the away-team jerseys are grouped into a single class and home-team jerseys are grouped in classes according to their jersey color. A convolutional neural network is then trained on the team identification dataset. The team identification network gets an accuracy of 97% on the test set. A novel player identification model is introduced that utilizes a temporal one-dimensional convolutional network to identify players from player bounding box sequences. The player identification model further takes advantage of the available NHL game roster data to obtain a player identification accuracy of 83%.

Motivation & Objective

• To develop an end-to-end automated system for player tracking and identification in ice hockey broadcast videos, where players are visually nearly identical except for jersey numbers.
• To address the challenge of high-speed, non-linear motion and dynamic camera movements in hockey videos, which complicate tracking and re-identification.
• To create and annotate new datasets for player tracking, team identification, and player identification due to the absence of publicly available benchmarks.
• To improve player identification accuracy by integrating game roster data and temporal context from player tracklets.
• To reduce identity switches and misclassification errors through post-processing and majority voting on probability windows.

Proposed method

• Player tracking is performed using state-of-the-art tracking algorithms, achieving a MOTA of 94.5% on the custom dataset.
• Team identification uses a convolutional neural network (CNN) that classifies players into home or away teams, with home teams clustered by jersey color and away teams grouped into a single class.
• A novel temporal 1D CNN model processes sequences of player bounding boxes (tracklets) to identify jersey numbers, leveraging temporal context across frames.
• The player identification model fuses tracklet features with game roster data by computing player-specific vectors (v_a and v_p) to refine predictions.
• An inference pipeline applies a visibility filter to detect frames where jersey numbers are likely visible, followed by probability aggregation using weighted averaging and post-processing to suppress false positives.
• Majority voting and null-class handling are used in post-processing to improve robustness, with the final prediction based on aggregated probabilities.

Experimental results

Research questions

• RQ1 How can player tracking and identification be effectively performed in fast-paced, camera-motion-heavy ice hockey broadcasts with minimal visual distinction between teammates?
• RQ2 What is the optimal method for combining temporal context from player tracklets with external game data (e.g., rosters) to improve identification accuracy?
• RQ3 How does visibility filtering and probability aggregation affect the robustness and accuracy of player identification in real-world broadcast video?
• RQ4 To what extent can team identification errors be mitigated through post-processing and majority voting on tracklet windows?
• RQ5 How much does incorporating game roster data improve overall player identification performance in a holistic tracking pipeline?

Key findings

• The player tracking component achieves a Multi-Object Tracking Accuracy (MOTA) of 94.5%, demonstrating strong performance on the custom dataset.
• The team identification model achieves 96.6% accuracy on the test set, with no need for fine-tuning on new games.
• The proposed 1D CNN-based player identification model achieves 83.17% accuracy with visibility filtering and post-processing, outperforming prior work by 9.9%.
• Incorporating game roster data increases the overall system accuracy from 77.9% to 82.8% on average across 13 test videos.
• For video 11, the accuracy improvement due to roster vectors is 24.44%, highlighting the significant impact of contextual data in low-visibility scenarios.
• The ablation study confirms that visibility filtering and post-processing are essential, as removing them reduces accuracy by over 2.37% even with optimal aggregation.