Infrastructure construction in cold regions requires year-round operations to prevent project delays, making winter construction an essential component. However, prolonged exposure to low temperatures can harm workers" health and compromise construction quality. Traditional monitoring methods, including wearable device detection and questionnaire surveys, face bottlenecks such as high costs, operational disruptions, and poor real-time performance. To address these challenges, this paper proposes a computer vision based non-contact cold injury monitoring framework. Utilizing 33 skeletal keypoints extracted by MediaPipe as input, it constructs two distinct methods: one based on deep learning and another on action geometry thresholds, both designed to identify characteristic warming behaviors. The paper systematically compares the characteristics of these two approaches in terms of accuracy and computational overhead. The geometric thresholding method was ultimately selected for experimental validation in Heilongjiang Province, monitoring three typical tasks: snow shoveling, material handling, and bricklaying. Results demonstrated 99.7% accuracy in identifying warming actions, confirming that the selected three warming actions effectively reflect cold injury status. This approach provides an efficient, low-cost technical solution for cold injury assessment in low-temperature construction scenarios.