Evaluation of Oil Workers' Performance Based on Surveillance Video

We present our research on the applicability of computer vision techniques for extracting various oil workers' performance metrics. This paper focuses on learning two metrics associated with the workers' location. The first metric e1 is the percent of frames in which only some part of the crew is present. If its value is bigger than some threshold value, the crew's performance is declared inefficient. We propose to perform human detection in each video frame and count people present in order to calculate e1. The Faster R-CNN and single-shot detectors with several types of feature extractors were tested on a specially collected dataset. By finetuning the most accurate of them we've achieved 0.99 precision and 0.91 recall. The second metric e2 considers workers' distance from an automated gas control system, which is the main subject of maintenance. We propose using some markers on the uniform for worker recognition and estimation of his/her position relative to an automated gas control system. We've tested the ArUco and the RUNETag markers on synthetic data and proved that they cannot be applied to our problem. We've also carried out some preliminary research on uniform numbers detection, as they can be also considered as markers. The Connectionist Text Proposal Network (CTPN) used for text detection achieved an accuracy of 0.76. Text recognition performed by Tesseract OCR failed with 0.05 recall. However, we plan to collect a dataset for number detection and recognition in the future and test more approaches.