Pedestrian detection method based on improved region regression

Abstract

The invention discloses a pedestrian detection method based on improved region regression, which comprises marking training samples; a region generation network including convolution feature extraction, semantic segmentation layer, classification layer and region regression layer being constructed, and its parameters being updated iteratively by training samples; a deep convolution neural networkincluding convolution feature extraction semantic segmentation layer and classification layer being constructed and its parameters being updated iteratively by using the candidate regions obtained inthe previous step; acquiring an image of a pedestrian to be detected; the image to be detected being inputted into the trained region generation network, and a plurality of candidate regions and confidence scores of each candidate region p predicted as pedestrians and background being obtained. The multiple candidate regions obtained in the previous step are selected to obtain confidence scores ofeach candidate region q predicted as pedestrian and background by using the trained depth convolution neural network with the first Ntop inputs. The probability that the candidate region q is predicted to be a pedestrian is obtained by combining the results of the first two steps. This method can provide an end-to-end pedestrian detection scheme by fusing multiple network outputs.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a pedestrian detection technology based on deep learning. Background technique [0002] Pedestrian detection, as a main branch of target detection, mainly detects and recognizes pedestrians in various postures. Whether it is the monitoring security warning system in public places or the automatic driving technology of automobiles, it has very common application scenarios. Various algorithms and framework systems for pedestrian detection have also emerged in recent years. Traditional machine learning methods use extracted image features such as integrated channel features (ICF), RotatedFilters and Checkerboard to detect pedestrians. In addition, deep learning methods based on neural networks are also applied to pedestrian detection. End-to-end pedestrian detection can be achieved by using convolutional neural network (CNN). Thanks to GPU acceleration technology...

AI Technical Summary

Problems solved by technology

[0004] (1) For machine learning methods that rely on the extraction of image features, they are not only affected by pedestrian shapes, angles, and similar distractors, but also sensitive to the training set, which is easily affected by the "noise" of wrong labels

[0005] (2) The problem of crowd mutual occlusion in multi-person scenes is prominent, but at present, most pedestrian detection algorithms do not deal with such occlusion situations specially, and there are few optimizations for such problems. Pedestrian position deviation or positioning error will occur, which will affect the accuracy of pedestrian detection

[0006] (3) The influence of non-pedestrian distractors, including the occlusion of foreground objects and pedestrian-like objects are also common in actual scenes. When the disturbers are relatively close to pedestrians, the existing methods rely on color and other features to detect the effect. good

Images