Post-Processing Algorithm for Workplace Safety Risk Monitoring System Based on Deep-Learning Models

Surveillance cameras are ubiquitously used as a part of the safety risk monitoring system. Powered by the artificial intelligence (AI) technologies, the system started to detect the anomalies by implementing deep-learning detection models. However, missed or unexpected warnings are happening frequently. And thus this problem is considered as a major flaw to the AI safety risk monitoring system. This study proposes a post-processing algorithm called PDAR aiming to enhance the accuracy of workplace safety risk monitoring system powered by deep-learning models. The PDAR consists of 3 types of algorithms with 3 typical situations: making warning when an action or an object detected, making warning when an absence or an inaction detected and making warning when joint objects or actions detected. The warning accuracy is significantly increased with PDAR validated with intensive experiments. An improvement of warning accuracy from 55.15% to 98.53% was observed in the test result. Moreover, an evaluation framework is also introduced which evaluate the warning accuracy by considering the missed warning and unexpected warning equally since either a missed warning or an unexpected warning would make serious consequences in safety risk monitoring scenarios. The dataset used in the experiments are now open sourced at: www.github.com/freelectry/PDAR.