Traffic Police Gesture Recognition
8-class pose classification from OpenPose skeletal features
Python OpenPose scikit-learn OpenCV · Jun.–Aug. 2020 · USC INF 552
An end-to-end 8-class gesture-recognition pipeline. OpenPose BODY_25 keypoints are transformed into 48-dimensional scale- and position-normalized features over a verified 323-sample modeling dataset.
MLP, ExtraTrees, NuSVC, and HistGradientBoosting were compared across ten randomized 60/40 splits; ExtraTrees reached ~92.4% mean and 95.3% best test accuracy. Normalized confusion matrices drove the error analysis — class-level errors, missing keypoints, class imbalance — and a Tkinter demo handles image selection, skeleton display, and four-model prediction.