Monocular far infrared pedestrian detection method based on feature fusion

Abstract

The invention discloses a monocular far-infrared pedestrian detection method based on feature fusion, and the method comprises the following steps of zooming an original infrared image according to different proportions, and obtaining a corresponding zoomed infrared image; carrying out local adaptive double-threshold segmentation binaryzation on the scaled infrared image; carrying out morphological processing on the obtained binary image; searching the processed binary image, and filtering the binary image to obtain a preliminary candidate region (ROIs); carrying out window sliding on the preliminary ROIs to obtain a series of sliding window ROIs; calculating the feature values of the ROIs, carrying out the normalization processing, and then inputting the feature values into a first SVM classifier and a second SVM classifier to be cascaded for decision making, and obtaining a preliminary pedestrian detection box; calling a non-maximum suppression algorithm to calculate the preliminarypedestrian detection box to obtain a final pedestrian detection box. The method can adapt to different distance detection, overcomes the pedestrian image breakage condition, and enables the pedestriandetection accuracy to be improved.

Description

technical field [0001] The invention relates to the research field of pedestrian detection, in particular to a monocular far-infrared pedestrian detection method based on feature fusion. Background technique [0002] At present, a vehicle-mounted far-infrared pedestrian detection system and method based on local features has been disclosed (Liu Qiong, Wang Guohua, Shen Minmin. Vehicle-mounted far-infrared pedestrian detection system and method based on local features: CN2014.), after local adaptive double threshold segmentation And morphological processing searches for ROIs from infrared images, uses a three-level cascade classifier to detect pedestrians in ROIs, and performs nearest neighbor matching with candidate areas through multi-frame verification screening results to fill in missed pedestrians. The local adaptive double-threshold segmentation algorithm used calculates a low threshold and a high threshold for each pixel by using horizontal scanning lines according to ...

AI Technical Summary

Problems solved by technology

However, due to the fixed width selected by the algorithm, this method cannot be well adapted to pedestrians

Moreover, this method ignores the first 12 pixels and the last 12 pixels of each line of the infrared image, resulting in the inability to detect pedestrians within these ranges of the infrared image

Furthermore, after local adaptive threshold segmentation and morphological processing, when the brightness of the pedestrian in the original infrared image is not much different from that of the background, the pedestrian will be broken, resulting in a certain number of ROIs that do not completely contain the pedestrian. It not only reduces the efficiency of classifier decision-making, but also reduces the accuracy of pedestrian detection; a video surveillance system implementation based on OpenCV has been disclosed (Du Xiaofeng. Research and implementation of pedestrian detection and monitoring system based on OpenCV [J]. Electronic production, 2018 (09): 58-59+9.), studied the performance of pedestrian detection through HOG feature extraction and SVM classifier based on HOG feature, and the paper showed that the SVM classifier based on HOG feature has an Higher efficiency and robustness; as a descriptor, HOG features have a good description effect on edge direction distribution, can well distinguish pedestrians from other backgrounds, and have good illumination invariance, making It is widely used in the field of pedestrian detection

However, due to the variability of pedestrians, such as pedestrian posture, pedestrian clothing, pedestrian occlusion, etc., and the complexity of the background, such as the impact of bicycles on cyclist detection, etc., HOG features still have certain limitations. still needs to be improved

Images