Automatic Formation Recognition in Handball Using Template Matching

Spatial organization of players in invasion games, i.e., formation is an important tactical feature. In handball defensive formations mainly include 6:0, 5:1, and 3:2:1 with different numbers of players acting on the defensive lines. The automatic detection of such formations is necessary for the analysis of large datasets. Position data of the players allow for algorithm-based approaches, like template matching. Therefore, the purpose of this study was to evaluate the quality of formation recognition in handball using template matching. 20 matches of handball videos were analyzed by domain experts. The respective start, end, and defensive formation was annotated. The corresponding position data was extracted, aggregated and compared to ground truth templates. Solving the linear assignment problem, the formation with the most similar template was predicted as formation. Precision and recall were calculated and compare to the majority class baseline (assigning all formations to 6:0). A total of 1,548 formations were analyzed in the final dataset. Precision was 0.95, 0.60, and 0.52 for 6:0, 5:1, and 3:2:1, respectively. Recall was 0.97, 0.45, and 0.50 for 6:0, 5:1, and 3:2:1, respectively. Classification with template matching was therefore more accurate then themajority class baseline (accuracy= 0.85). Template matching seems to be a viable solution to classify between space-oriented (6:0) and more man-oriented (5:1, 3:2:1) formations but less able to identify differences between man-oriented formations. Future work can include more information than the average player position, e.g., the players' covariance matrices to get more information about role specific movement behavior.