Semi-supervised pedestrian detection method based on deep convolutional network

Abstract

The invention discloses a semi-supervised pedestrian detection method based on a deep convolutional network, and the method comprises the steps: randomly extracting 5% from a 100% complete training set of a public data set to serve as a training set of semi-supervised learning, reserving the annotations of 5% of the training set, taking the annotations of the 5% of the training set as annotated training sets, taking the rest 95% of the training set as unannotated training sets, and deleting the annotations. An improved RPN convolutional neural network is adopted. The consistency constraint regularization term is added to improve the feature expression capability of the network, so that the robustness of the model is improved; through self-paced learning, the dependency of the deep learningnetwork on a large number of manually labeled samples is reduced, and the over-fitting risk of the deep learning network under the condition of small samples is reduced.

Description

technical field [0001] The invention relates to the field of pedestrian detection, in particular to a semi-supervised pedestrian detection method based on a deep convolutional network. Background technique [0002] Pedestrian detection is an important topic in the field of computer vision. Pedestrian detection in the perception of scene information is a crucial part in the application of unmanned driving and video surveillance. Unmanned driving requires computers to be able to accurately identify pedestrians and obstacles to assist other intelligent driving technologies to predict road conditions and reduce traffic accidents. At present, unmanned driving technology is favored by many companies such as Google and Baidu, and a large number of scientists are actively developing it. Shopping malls, transportation and other public places are equipped with surveillance to maintain good public order. Manual processing of surveillance video information requires a lot of manpower an...

AI Technical Summary

Problems solved by technology

However, the existing deep learning model pedestrian detection method is very dependent on the size of the training data, which poses a big challenge: for a good model between a small number of labeled samples and a messy background, the labeled samples are not enough. It is easy to misjudge many false positives when the shooting angle is tricky

And unlike public datasets that have good shooting angles and enough pedestrian samples to learn rich target information, most video surveillance does not have enough annotations, and low pixels, small pedestrian target scales lead to small intra-class differences, which seriously degrades the detection performance

Images