Taekwondo is a martial art from South Korea that is gaining popularity in various international competitions. Judging in Taekwondo is often done through scoring by referees, sensors, or cameras to get objective and accurate results. This research focuses on developing a Taekwondo kick evaluation system using Convolutional Neural Network (CNN) with ResNet50 visual feature extraction. We compared the effectiveness of ResNet50 in detecting normal and anomalous kicks with K-Nearest Neighbor (KNN) and Support Vector Machine (SVM). The dataset used comes from the open access Pose Sequence Score (PSS) sensor. A data augmentation method was applied to increase the diversity of the training data, and the ResNet50 model was modified to maximize performance. The results showed that CNN-ResNet50 achieved 88% accuracy with balanced precision, recall, and F1-score. However, the KNN and SVM methods provided higher performance with 100% and 99% accuracy, respectively. Although CNN-ResNet50 is quite effective, KNN and SVM proved to be a better choice for Taekwondo kick move classification.