Study Of Channel Combinations On ACF Algorithm

Pedestrian detection is an important research content in the field of computer vision.It extracts basic semantic information in video clips,and has wide application prospects in areas such as automatic driving,intelligent monitoring,pedestrian analysis,and intelligent robots.In the past ten years,pedestrian detection technology has developed rapidly,and a large number of pedestrian detection methods have been proposed.Pedestrian detection has greatly improved the recognition accuracy and recognition speed.In 2009,Dollar et al.proposed the integral channel feature,used multiple feature channel fusions to extract information,and achieved more advanced results at that time.The ACF,aggregated channel features,proposed to improve the integral channel feature is the best pedestrian detection effect.One of the methods.On the one hand,aggregated channel features using feature pyramid and pooling simplify the feature calculation and further improve the detection effect.On the other hand,there is still room for further exploration and improvement of the characteristics of the aggregated channels: the understanding of the role of each channel and the exploration of the combined effect of each channel is still insufficient;the three types of 10 feature channels used may be unnecessary Redundant information still has room for simplification;for different application scenarios,different combinations of feature channels can still be used to determine the speed or accuracy of pedestrian detection.This article focuses on the in-depth experimental exploration of the aggregated channel features.Firstly,the development of pedestrian detection,current research status,current mainstream detection methods,basic knowledge of pedestrian detection,and the principles of integral channel characteristics and aggregate channel characteristics are introduced.Then,a combination experiment is performed on each feature channel of the ACF algorithm,including experiments for each channel to work independently,for each combination,and for changing the number of channels in the gradient orients.Through experiments,the effects of each channel are analyzed,and several simplified channel combinations with practical meanings are found.This paper proposes that in different practical application scenarios,different simplified combinations of the original 10 feature channels of the aggregated channel feature can be performed to make a trade-off between the accuracy and speed of pedestrian detection and to better adapt to pedestrians in different scenarios.