Design and Implementation of a Monocular Vision-Based Door Handle Grip Point Localization Model

In real life, doors are ubiquitous, and opening doors has become a “necessary skill” for general-purpose robots. The key for robots to achieve the task of opening doors is to obtain accurate three-dimensional spatial coordinates of the grip point of the door handle. The commonly used binocular camera localization technique gets the target depth information through parallax, but there are defects such as visual blindness. To recognize the door handle in real-time and predict the 3D coordinates of the door handle grip point, this paper adopts a monocular camera localization technique combined with an advanced target detection algorithm (yolov5) to design a monocular vision-based door handle grip point localization network: a total of target detection module and target localization module are included. The target detection model incorporates BiFPN feature fusion to extract effective features and fuse them through a bidirectional pyramid structure; the SMOTE overfitting sampling algorithm is incorporated to compensate for the uneven sample distribution by generating some small samples. The experimental results show that the detection accuracy and classification accuracy of the model in this paper are higher than those of yolov5 and other models.