A Two-Step Classification Platform to Classify Construction Sounds to Enhance Real-Time Construction Zone Monitoring
Semester of Graduation
Master of Science in Computer Science (MSCS)
Computer Science and Engineering
In recent years, sounds generated from equipment used at a construction site have been found to be useful in identifying construction activities and associated equipment usage providing insightful information regarding project status and hazard issues. This study proposes a Two-Step Neural Classifier (TSNC) for accurate classification of real world construction sounds. Experimental results from real world construction site audio data showed an average classification accuracy of 98% for TSNC compared to 91% average accuracy for an ordinary (one-step) Artificial Neural Network (ANN). The TSNC model is a hierarchical model that exploits the idea of creating disjoint subgroups of construction sounds and uses separate ANN classifiers for each sub group. Since each subgroup ANN classifier only handles approximately one-third of the total number of sound-classes, each of those ANNs perform much better than a single ANN handling all sound classes. The subgroups of activities are decided through k-Means clustering. Certain construction sounds like Hammering, Truck(Dumper), Dozer and Welding sounds are found to have within them two or more linear clusters in the plot of average of absolute values of sound samples and their standard deviations, indicating that those construction-sounds are more complex and need to be treated as a mixture different sound classes instead of an atomic class. Including them as separate classes in the TSNC model caused the classification accuracy to go down indicating the need for additional study.
Jui, Jayati Halder, "A Two-Step Classification Platform to Classify Construction Sounds to Enhance Real-Time Construction Zone Monitoring" (2019). LSU Master's Theses. 4997.
Available for download on Wednesday, August 19, 2026