An improved target detection method based on multiscale features fusion

Abstract

AbstractA pedestrian‐target detection method based on multiscale convolution features is proposed to solve the problem of the existing target detection algorithms based on convolutional neural networks that cannot effectively adapt to the target scale change, its deformation, and the complex environment. This article analyzes the characteristics of small target, establishes the a three‐layer pyramid network structure based on horizontal connection fusion, and makes full use of shallow convolution features with lower semantics to improve the accuracy of small target detection; From the angle of the cross ratio, quantitative analysis is made on the effect of eliminating redundant bounding box by nonmaximum suppression based on center point when the window with the cross ratio is less than 0.7, solves the problem of missing detection of overlap and small target, and improves the target detection accuracy and the generalization ability of the target detection model. Through the experimental results on VOC2007 and VOC2012 public data sets, the proposed method of this article has higher detection accuracy and stability for small and incomplete targets, multiscale target detection, and target classification under the complex environment.