In the construction industry, fall accidents are the leading cause of construction-related fatalities; in particular, ironworkers have the highest risk of fatal accidents. Detecting near-miss accidents for ironworkers provides crucial information for interrupting and preventing the precursors of fall accidents while simultaneously addressing the problem of sparse accident data for ironworkers' fall-risk assessments. However, current methods for detecting near-miss accidents are based upon workers' self-reporting, which introduces variability to the collected data. This paper aims to present a method that uses Inertial Measurement Unit (IMU) sensor data to automatically detect near-miss accidents during ironworkers' walking motion. Then, using a Primal Laplacian Support Vector Machine, a developed semi-supervised algorithm trains a system to predict near-miss incidents using this data. The accuracy of this semi-supervised algorithm was measured with different metrics to assess the impact of the automated near-miss incident detection in construction worksites. The experimental validation of the algorithm indicates that near-miss incidents may be estimated and classified with considerable accuracy-above 98 percent. Then the computational burden of the proposed algorithm was compared with a One-Class Support Vector Machine (OC-SVM). Based upon the proposed detection approach, high-risk actions in the construction site can be detected efficiently, and steps towards reducing or eliminating them may be taken.