Research On Pedestrian Detection Based On General Object Detection And Haar-like Characteristics

As an important topic in the field of computer vision,pattern recognition attracts a large number of experts and scholars.Pedestrian detection is an important branch of pattern recognition and plays an increasingly important role in intelligent monitoring,au-tomatic driving,and intelligent robots.This paper focuses on the research of pedestrian detection in static images.For some limitations of the current existing pedestrian detection algorithms,this pa-per proposes some improvements:Different from the traditional "one-step" approach to pedestrian detection,this paper proposes to divide the pedestrian detection into two parts.First,the BING feature is applied to pedestrian detection.By binarizing the approximate NG feature,the time spent on feature extraction is significantly reduced.We improved the original linear model of the BING algorithm to highlight the relevant characteristics of pedestrians in preparation for the subsequent improvement of pedestrian detection accu-racy.Through this step detection,approximately 400 candidate boxes are obtained.This step screens candidate boxes without object features and reduces subsequent unnecessary calculations.With this step improvement,the screening time for candidate frames can be reduced by three orders of magnitude.Second,the candidate frame obtained in the first part is sent to the detector for detection,which distinguishes the pedestrian from the ob-ject and achieves the purpose of pedestrian detection.We selected 10 feature channels for training and testing,including the gradient,LUV,and HOG channels.In terms of feature pool construction,we propose a Haar-like feature based on the visual three-part features of the pedestrian(head,upper half,and lower half),and construct a suitable feature pool through the weight maps of each part of the human body.It is more suitable for pedestrian detection in this specific scenario.With this improvement,we can improve the detection accuracy under the same conditions.The experimental results show that Our improvements can significantly reduce com-putation time complexity and space complexity,while the detection accuracy has a good performance on the INRIA and Caltech data sets.