Pedestrian Detection Based On Deep Learning

Pedestrian detection technology has always been an important research direction of computer vision.In recent years,with the practice of pedestrian tracking,gesture recognition,gait analysis and other technologies gradually applied to daily life,pedestrian detection has been increasingly researched and applied.With the continuous development of detection algorithms,pedestrian detection technology has also evolved from the early model extraction feature with model classification phase to the deep learning phase.Although the pedestrian detection based on deep learning has achieved good results.But in practice applications,detection performance is significantly reduced because of pedestrians are different in size,environmental factors are also complex and variable.The detection performance under non-ideal conditions tends to decrease significantly.For the problem of small target pedestrians' missed detection rate and high false detection rate,introducing Tiny-YOLOv2 convolutional neural network,and the above small target problem is improved.The specific research contents are as follows:Firstly,in view of the long and thin characteristics of the pedestrian bounding box,in order to make the target bounding box of the detection algorithm more in line with the pedestrian length and width ratio,using K-means++ clustering algorithm to cluster the proportion of the pedestrian bounding box,and the YOLO network framework is re-planned.The number of lower target bounding boxes is proportional to the aspect ratio of the target bounding box.Improve the fit of the target bounding box to the pedestrian and the speed of the network calculation.Secondly,based on the Tiny-YOLOv2 convolutional neural network,proposing a small feature fusion method for small target pedestrians easily lose information and locate inaccurate.The feature fusion calculation based on YOLO network framework is proposed.Introducing the idea of multi-scale feature fusion,selecting the feature maps of different layers for superposition in the calculation of the final feature map,combining shallow detail features and deep semantic features to reduce the loss of small target pedestrians information caused by continuous feature extraction and pooling process,improving the characteristic expression of small target pedestrians.Finally,in order to evaluate the commonly used deep learning network model,a comparative experiment was conducted on networks in the field of pedestrian detection,such as YOLOv3,Faster R-CNN and SSD,in the same experimental environment.The INRIA and PASCAL VOC data sets were used as test sets respectively.The characteristics of three deep learning network models in pedestrian detection were analyzed from the aspects of accuracy,detection speed and detection ability of target small pedestrians.The experimental results show that after cluster analysis,the coincidence degree between the pedestrian bounding box and the anchor boxes is significantly increased,and the time required to train the network model is shorter.Tiny-YOLOv2 network after feature fusion have higher detection accuracy and recall,in the small target pedestrian dataset are more obvious.