An Object Detection Method Based on Bounding Box Corner Alignment and Boundary Matching

Abstract

The invention discloses a target detection method based on bounding box corner point alignment and boundary matching, comprising the following steps: (1) acquiring a real scene picture, marking the category of the target object and the position of the bounding box to form a training data set; (2) Input the pictures in the training data set into the detection model to obtain the predicted category distribution and predicted bounding box position of the objects in the picture; (3) construct a loss function to calculate the classification loss of the object and the positioning loss of the bounding box respectively; (4) optimize the above classification The loss function of loss and positioning loss, select the pictures in the training data set, repeat steps (2) and (3), and end the training after reaching the preset training times; (5) After the detection model is trained, select the picture to be detected and input it into the model, Get the category and bounding box location of the object. With the present invention, the network can be made to focus on the learning of bounding boxes with a relatively low degree of coincidence, and the overall detection accuracy can be improved.

Description

technical field [0001] The invention belongs to the field of computer vision target detection, in particular to a target detection method based on corner point alignment and boundary matching of a bounding box. Background technique [0002] Object detection is an important task in the field of computer vision. It is a field that promotes and develops with deep learning, and can be applied to real-world scenarios such as vehicle detection, pedestrian detection, traffic light detection, as well as driverless, security systems and many other fields. In recent years, with the development of convolutional neural networks and the introduction of various deep learning models well-designed for object detection tasks, the problem of object detection has achieved amazing progress. [0003] For example, the Chinese patent document with publication number CN111428625A discloses a deep learning-based traffic scene target detection method and system. The improved YOLOv3 target detection ...

AI Technical Summary

Problems solved by technology

[0005] However, when the IoU-based loss is relatively small in bounding box coincidence, the resulting gradient will be relatively small.

Therefore, in the process of training the neural network, the gradients generated by some bounding boxes with relatively large coincidence degrees will dominate the gradients of the backpropagation of the neural network, so the gradients generated by some bounding boxes with relatively small coincidence degrees are easy to be ignored, affecting neural networks. Network Learning on Difficult Examples

Therefore, the overall final detection accuracy is not high

Images