Serial channel fusion pedestrian detection method based on binocular vision

Abstract

The invention discloses a serial channel fusion pedestrian detection method based on binocular vision. According to the method, a parallax image is generated according to a binocular RGB image, a convolutional neural network is constructed, an RGB image channel and a parallax image channel are connected in series to serve as input of the convolutional neural network, the complementary relation between the RGB image and the parallax image is learnt, a part with a predicted value greater than a set value is selected from a detection result output by the convolutional neural network, non-maximumsuppression is performed on the part, and a final pedestrian position is obtained. Through the method, since the RGB color image has rich color information and the parallax image is invariant to illumination, pedestrian detection can be more robust and accurate by combining the RGB image and the parallax image.

Description

technical field [0001] The invention relates to the research field of pedestrian detection methods, in particular to a binocular vision-based serial channel fusion pedestrian detection method. Background technique [0002] Pedestrian detection is to locate the position of pedestrians from a given image. People are the main body of society, so pedestrian detection is also the most common task among many target detection tasks. At the same time, pedestrian detection is a basic part of advanced visual tasks such as pedestrian tracking and pedestrian re-identification. The development of pedestrian detection methods is of great significance to the development of applications such as cross-camera pedestrian tracking and pedestrian search. For example, based on a good pedestrian detection result, we can track the position of the pedestrian. In addition, through the pedestrian re-identification technology, we can identify where the pedestrian appears in another camera, so as to re...

AI Technical Summary

Problems solved by technology

However, on the one hand, due to the weak robustness of RGB information to illumination and color changes, the pedestrian detection method based only on RGB images will face greater challenges in the case of drastic illumination changes and complex backgrounds.

On the other hand, although the edge or contour information extracted from the RGB image is robust to illumination and color changes, it is difficult to see some negative samples shaped like pedestrians (such as shadows shaped like pedestrians) and some Is it a positive sample of pedestrians (such as pedestrians with colorful clothes in complex backgrounds) or weak discrimination

The actual contour information of the object is more robust to this situation, but it is difficult to learn the actual contour information of the object from the RGB image, which will increase the learning pressure of the model

Images