Research On Pedestrian Detection Algorithm Based On Deep Learning

Pedestrian detection technology is an important research branch in the field of computer vision and target detection,which is widely used in artificial intelligence,vehicle assisted driving,intelligent monitoring,intelligent traffic and other aspects.Based on traditional machine learning methods for real-time performance,the pedestrian detection accuracy is low,this paper will deep learning combined with pedestrian detection technology,adopted multiple classification unipolar detector namely SSD algorithm as the basis of pedestrian detection method,and then on the SSD algorithm is improved,the two kinds of new improved algorithm,Experimental comparison in INRIA pedestrian detection data set shows that the improved algorithm has higher detection accuracy and shorter training time than the original SSD algorithm,indicating that the improved algorithm is more optimized in real-time and accuracy of pedestrian detection.First of all,this paper describes the research background and significance,investigates the research status at home and abroad,and analyzes the differences and better development prospects of pedestrian detection technology under traditional methods and deep learning methods.The basic knowledge of convolutional neural network,algorithm optimization method and evaluation index are introduced.Secondly,this paper introduces the SSD algorithm and its feature extraction network,multi-scale feature mapping,bounding box regression,non-maximum suppression,loss function and so on.INRIA pedestrian detection data set was self-made,and SSD algorithm was used to conduct experiments on the self-made data set,and the detection accuracy was72.94%.This paper also proposes an SSD algorithm based on residual structure.The VGG16 feature extraction network in the SSD algorithm is replaced by Res Net residual structure.Experiments are carried out on INRIA data set,and the experimental results are compared with those generated under the original SSD algorithm.And the training process is much more robust.Finally,in order to make the improved algorithm more stable and accurate in the training process,an SSD algorithm with deep residual shrinkage structure was proposed,and the classification network in the algorithm was replaced by the deep residual shrinkage structure DRSN-CW.The deep residual structure,the principle of deep residual shrinkage structure and the network structure of the improved algorithm are introduced.The detection accuracy of the further improved algorithm is 80.13% on the pedestrian detection data set.By comparing the loss function line graph and detection accuracy generated by the above three algorithms in the training process,it is concluded that the SSD algorithm with deep residual shrinkage structure has the best real-time performance and accuracy among the three networks.