People Detection in an Elevator Car Using Computer Vision

There is a need to develop an integrated sensor system in elevator cars to cover as many use cases as possible to address the increasingly stringent laws implemented by governing bodies to ensure passenger safety. This research focused on people detection using computer vision as the first step towards a complete technological solution. Six-phase design methodology was applied. An embedded device was developed with Nvidia Jetson Nano (2GB) and 200° field of view camera. Dataset of 267 images of passengers in real elevator setting was obtained using the device. The performance of existing object detection algorithms - YOLOv5, MobileNetV2SSD and Roboflow AutoML were evaluated by re-training the models with captured image dataset and barrel-corrected image dataset. The performance of YOLOv5 and Roboflow AutoML with barrel-corrected image dataset was similar with promising mean average precision (mAP) whereas mAP for MobileNetV2SSD was subpar. Further improvements should be made to eliminate ghost detections and spurious detections caused due to varying elevator environments with highly reflective surfaces.