Based On Combining Model For Pedestrian Detection

Pedestrian detection is always a high-profile problem in computer vision,and there are a lot of applications in real life need to use pedestrian detection technology,for example smart driving,video surveillance and so on.Although the research of pedestrian detection started very early,however,there are still many problems to be solved in the field of pedestrian detection,such as pedestrian gesture,pedestrian size in the picture,objects obscure pedestrians,something shaped like a pedestrian and so on.All these factors we talked above can affect the result of pedestrian dection,and because of this,these matters are also the more important research directions in the field of pedestrian detection.This paper uses a statistical learning method to conduct pedestrian detection tasks,and the method of extract features is CNN-based method of deep learning filed.In this paper,a new pedestrian detection model has been proposed,it combines the method of extract features with CNN-based method of deep learning filed and the method of classification with decision tree of traditional classification method together.Finally,the model we proposed achieves rapid and accurate detection of pedestrians.In particular,the detection effect of small targets has been greatly improved.The main research contents of this work are as follows:(1)Because of the quality of the features can directly affect the detection accuracy of the whole pedestrian detection model,so this paper analyzes and studies the feature maps generated by different convolution layers in convolutional neural networks,and through the combination of features to get the feature vector,then through the classifier to carry out classification tasks.Finally,experiments show that the features generated in this way have good recognition ability and greatly improve the recognition accuracy of the pedestrian detection model.(2)This paper adopt a new classifier called XGBoost,and it performs training tasks by used outputs of region proposal network and get the final result.Compared with the classifier based on the convolutional neural network method,XGBoost classifier can achieve zero loss of use of input information,and classifier training and testing tasks can be carried out in parallel on the CPU.The process does not rely on the GPU at all,but still has a very fast detection speed.(3)The experimental data used in this paper is the very popular in the field of pedestrian detection called Caltech Pedestrian database.The proposed method can reduce the missed detection rate to 8.77% under the "Reasonable" testing standard,And in terms of speed of detection,it is ahead of other advanced methods in pedestrian detection.When not considering the issue of detection speed,by adding global standardization and optical flow method two ways,detection accuracy can be further improved to 8.25% of the missed detection rate.