The application of human activity recognition based on motion sensors in sanitation work offers significant potential for real-time monitoring of worker activities and identifying potential safety hazards. This paper explores the use of machine learning techniques to recognize and classify seven distinct types of sanitation worker activities: walking, running, sweeping, sweeping with a big broom, cleaning, dumping, and daily activities (such as sitting and smoking). Data were collected using a triaxial accelerometer in a wrist smartwatch, sampled at 25 Hz, resulting in a comprehensive dataset of 266,555 samples. Preprocessing involved segmenting the time series data into 5,026 windows and extracting 57 time-domain and frequency-domain features for each window. Various machine learning classifiers, including Decision Trees, k-Nearest Neighbors (KNN), Neural Networks, Support Vector Machines (SVM), and Naive Bayes, were employed to analyze the data. The Subspace KNN classifier achieved the highest test accuracy of 93.8% using the Relief feature ranking algorithm on the time-domain dataset with 30 features and an 80% training rate. In contrast, the Bagged Trees classifier achieved the highest test accuracy of 84.9% using the MRMR feature ranking algorithm on the frequency-domain dataset with 15 features and a 90% training rate. These results demonstrate the effectiveness of machine learning in enhancing worker safety and monitoring in sanitation activities.