A Scale-Matching-Based Detection Method for Weak People

Abstract

The present invention provides a method for detecting weak and small person objects based on scale matching. The method includes the step of transferring the scale distribution of persons on the network pre-training data set to the scale distribution of weak and small persons on the target training set; when training the object detection model, first Pre-training on the scale-shifted pre-training data set to obtain a preliminary object detection model, and then train on the target training set to obtain the final object detection model. Through scale matching, the method of the invention enables the model to better study and utilize information on a small scale, makes the convolutional neural network or other models more accurate in the representation of weak and small targets, and effectively improves the detection performance of the prior art detector.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to a scale-matching-based weak and small person target detection method. Background technique [0002] Pedestrian detection is an important topic in the field of computer vision and has a wide range of applications, including surveillance, driver assistance, mobile robotics, and rapid rescue at sea. With the rise of deep convolutional neural networks, unprecedented progress has been made in pedestrian detection, however, the detection of weak people is far from good results. Unlike objects of proper scale, the detection task of tiny objects is more challenging due to their very small relative and absolute sizes and low signal-to-noise ratio. If the video is from the image in the video, after the video encoding and decoding process, the blurred image will cause tiny objects to be mixed with the background, which makes it more difficult to obtain tiny targets. In...

AI Technical Summary

Problems solved by technology

[0005] For the weak and small person target detection problem we want to solve, there is currently no public external data set that can meet most of the standards for weak and small people (such as weak and small people with a scale of less than 20 pixels). Therefore, in the target data set and The problem of different scale distributions on the pre-training data set will become the main factor that limits the performance of the pre-training model on the data set.

[0006] The purpose of the present invention is to provide a simple and effective detection method for weak and small people to solve the above-mentioned problems in the detection of weak and small people, such as the lack of target data sets related to weak and small people, and the current public pre-training data sets are not suitable for weak and small people. The problem of limited performance improvement in detection

Images