Mentor: Soyean Kim, BC Safety Authority
Summary
BC Safety Authority’s accident prevention strategy follows a risk based platform, called “bow tie” for ongoing management and prevention of major incidents. The bow tie method draws on the direct experience of safety management team and the data collected from various safety oversight programs in order to identify hazards and to properly incorporate critical controls into management systems. In addition, it raises awareness and improves understanding and risk knowledge of the potential major incidents and the reliance on critical controls that prevent those accidents from occurring. Opportunities exist for greater connectivity among various databases to maximize the risk knowledge, especially the pre incident data based on inspections and post incident data.
In the inspection process and incident reporting process, photo data of on-premise materials are recorded, and assigned a hazard rating level. As above, this hazard rating level contributes to resource allocation for risk mitigation. Automatic detection of hazardous objects in these photos could be used to bolster these efforts, and streamline inspector and analyst workflow alike.
To this end, there now exist state of the art image recognition algorithms, capable of reliably detecting objects of more than 1000 different object classes. We believe it is possible to use transfer learning as a starting point, in conjunction with thousands of labelled image data, to create an accurate classifier of hazard rating level.
Because of the nature of the images, in order to this effectively it might first be necessary to segment the hazardous object in the image. As no such labelling presently exists, and labelling thousands of images by hand would be very time consuming, it is suggested to use active learning methods to efficiently explore the feature space, thereby improving training time.