Video target detecting method based on cascade regression convolutional neural network

Abstract

The invention provides a video target detecting method based on a cascade regression convolutional neural network. The video target detecting method comprises the following steps of 1, inputting a video image sequence, and performing CNN characteristic extraction on all image frames of the whole video sequence through the convolutional neural network; 2, classifying the last convolutional characteristic layer of the CNN characteristic by means of an RPN network for obtaining a suggested area, performing cascade classification and regression on the suggested area through a multiscale convolutional characteristic, and obtaining a static picture detecting result of each image frame; 3, using the results with confidences which are larger than 0.6 in the detecting results that are obtained in the step 2 as initial tracking values, tracking the target on the conv5-3 convolutional characteristic of the CNN characteristic through related filtering for obtaining a time sequence suggested area,performing cascade classification and regression on the time sequence suggested area, and obtaining a time sequence detecting result; and 4, suppressing abnormal values in the static picture detectingresults and the time sequence detecting results through a co-occurrence matrix, thereby obtaining a final detecting result.

Description

technical field [0001] The invention belongs to the field of image information technology processing, and in particular relates to a video target detection method based on a cascade regression convolutional neural network. Background technique [0002] Object recognition is a method of automatically locating objects in images. It is a basic problem in the field of computer vision and has applications in many fields, such as monitoring, human-computer interaction, and medical assistance. Early methods can effectively detect single-category objects in images, such as faces and pedestrians, through sliding windows or cascaded classifiers, but cannot detect multi-category objects. [0003] In recent years, thanks to the development of convolutional neural networks, multi-category object detection technology has been significantly improved. Among them, the region-based convolutional neural network consists of region proposal and region classification, and R-CNN transforms target...

AI Technical Summary

Problems solved by technology

First of all, the region (region) obtained by the RPN network in each frame image is not always reliable, because the sharp appearance change of the target in the video will affect the recall rate of the network

Secondly, the separate region classification cannot integrate the spatio-temporal information and context information in the video, which makes the target score of the video more blurred too low, which affects the performance of detection

Images