Research On Multi-scale Pedestrian Detection Technology In Complex Background

As an important part of the object detection field,pedestrian detection aims to apply related technologies to determine whether there is a pedestrian in the input image and mark it.Nowadays,it is widely used in various fields of life such as intelligent transportation systems,human pose estimation,and medical care.With the rapid development of deep learning,the current pedestrian detection algorithms have achieved good results in many public datasets,but pedestrian detection in complex backgrounds is affected by lighting changes,background textures,pedestrian poses,and variable scales and other factors.The current pedestrian detection technology based on deep learning still has many problems to be solved.This paper studies the following three aspects of multi-scale pedestrian detection based on complex backgrounds,aiming at the problems of uneven illumination,changing pedestrian scale and pose,and distance from the camera.1.Aiming at the problem that the actual application background is complex and the illumination change has a huge impact on the detection effect,a wavelet noise reduction image enhancement model based on Retinex is researched and constructed,and the input image is preprocessed.Based on the Retinex algorithm,the contrast of the image under the complex background is enhanced,the image to be processed is converted from RGB to HSV space,and then the brightness V is enhanced.Using the Retinex principle and Weighted Variational Model(WVM),the lightness V is decomposed into the reflection layer R and the illumination layer L,and the shadow enhancement is performed on the decomposed illumination layer.On the premise of ensuring that the bright areas are not overexposed,a better balance between bright parts,dark parts and noise is achieved.2.Aiming at the phenomenon that pedestrians have different poses and different distances from the camera,the pedestrian scale changes greatly,and long-distance pedestrian detection is difficult.In this paper,an improved algorithm based on channel enhancement for YOLOv3 multi-feature fusion is proposed.Compared with the detection effect of different feature enhancements,we choose to add the Squeeze-and-Excitation module after the residual module of the skeleton network to improve the mining ability of multi-scale pedestrian feature information,and perform Spatial Pyramid Pooling at the end of the network,strengthen the network feature extraction effect.Accumulate the processed low-level features and processed high-level features,fuse multi-layer feature information,and output in different features,The multi-scale prediction of feature pyramid fusion is used to deal with small-scale pedestrians at long distances,thereby realizing the availability of the detection model at the multi-scale level and improving its detection accuracy.Training on the Caltech dataset,and verify the effectiveness of the algorithm on the COCO and Caltech test sets.3.Aiming at the phenomenon that the width and height scales of the target pedestrians vary greatly,and the conventional anchor frame leads to the failure of prediction,an improved algorithm for cross-scale anchor prediction is proposed.By adjusting the distribution of Anchor boxes,the detection system can better adapt to extreme-scale pedestrians,which further improves the multi-scale detection effect.By calculating the Euclidean distance between the two center points of the predicted frame and the real frame,the intersection area relationship between the predicted sample and the real frame is directly obtained.Compared with directly using the mean square error loss to calculate the distance between the sample and the real frame,the improvement makes the training process more stable to complete the regression of the target frame,thereby avoiding the situation that it is not easy to converge.Finally,combining the image enhancement model and the pedestrian multi-scale feature fusion process,a multi-scale pedestrian detection system is constructed,which realizes the effectiveness of pedestrian detection in complex backgrounds.