Improving Industrial Safety Gear Detection through Re-ID conditioned Detector

Industrial safety gears such as hardhats, vests, gloves and goggles are vital in safety of workers. With the advancement of vision technologies, most industries are moving towards automatic safety monitoring systems for its enforcement. However, most of the industrial safety monitoring systems are plagued by the following problems. To begin with, object detection which is the principal component of this system suffers from the problem of false detections and missed detections which are extremely costly resulting in wrong safety monitoring alerts and safety hazards. Further, while video object detection has seen a large traction through ImagenetDet and MOT17Det challenges, to the best of our knowledge there is no work till date in the context of industrial safety. Finally, unlike existing areas of object detection where there is the availability of large datasets, best of existing research works in detecting industrial safety gears is restricted to mostly hardhats due to lack of large datasets. In this work, we address these previously mentioned challenges by presenting a unified industrial safety system. As part of this developed system, we firstly introduce safety gear detection dataset consisting of 5k images with the previously mentioned classes of safety gears and present exhaustive benchmark on state-of-the-art single frame object detection. Secondly, to address wrong/missed detections we propose to exploit temporal information from contiguous frames by conditioning the object detection in the current frame on results of re-identification of objects computed in prior frames. Finally, we conduct extensive experiments using the developed Re-ID conditioned object detection system with various state-of-the-art object detectors to show that the proposed system produces mAP of 85%, 87%, 92% and 78% with average improvements of 5% mAP across the previously mentioned safety gears under complex conditions of illumination, posture and occlusions.