Abstract

The fall accident is one of the major causes of fatal injuries and financial losses in the construction industry. Currently, many methods have been studied to detect fall accidents on the construction sites to avoid fall injuries and reduce financial losses. The state-of-art fall detection methods include wearable sensor based fall detection, ambience based fall detection and vision based fall detection. However, these methods still have limitations in terms of accuracy and detection speed, when being used to detect and locate fall accidents on the construction sites in practice.
The main objective of this research is to propose a novel vision based framework to detect and locate fall accidents on the construction sites promptly and automatically. In order to achieve the main objective, three methods (worker localization, worker matching, and fall detection) are created under the proposed framework. The worker localization method acquires real-world map coordinates from video frames based on the perspective transformation. The worker matching method matches workers captured by different camera views based on their spatial relationship according to the construction sites. The fall detection method employs an artificial neural network. The neural network is trained with features extracted from videos to detect fall accidents automatically.
Experiments have been conducted both in lab and on real construction sites to test the performances of the methods under the proposed framework. The experiment results indicated that the average localization accuracy was 90%. The accuracy is similar to the previous works; however, no attached sensors or tags are required with the proposed method. The matching accuracy was 93.01%. Compared with the method proposed by Lee et al. (2016), the proposed method is more accurate when cameras are set far from workers. The fall detection had an 83% precision and a 90% recall rate. The accuracy is similar to the previous works; however, the proposed method does not require subtle vision features of workers.
The main contribution of this research study is proposing a framework providing information about fall accidents on the construction sites promptly and automatically. Also, the methods created in this research study can be used to assist other automated construction processes including tracking, motion detection, etc. Future works will focus on improving the localization accuracy, matching workers under ultra-large baseline camera networks and implementing deep neural networks for fall detection.