Monitors in restaurants and retail stores
posted in 2021
posted in 2021
(I) Objective: Detect people in restaurants and retail stores.
(II) Approach: Due to the company's chip not supporting the Upsampling layer, the YOLOv5-s model structure was modified by removing the PANet, as shown in (Fig. 1).
(Fig. 1a) Original YOLOv5 model structure.
(Fig. 1b) Modified YOLOv5 model structure.
(III) Dataset: Trained on open public datasets with person class bounding boxes, including CrowdHuman, COCO, and PASCAL VOC.
(IV) Result: The model successfully detected people and heads. By pairing the person and head bounding boxes using Intersection over Union (IoU), false positive person detections were reduced. The below images show the modified YOLOv5-s model's predictions, including person bounding boxes (blue) and head bounding boxes (red).
The project succeeded because my model can accurately detect people.