Research On Pedestrian Detection In Complex Scene Based On Deep Learning And Migration Learning

In recent years,with the continuous development of hardware resources and computer vision,the combination of these is more popular in all walks of our life,and the computer vision is an important direction.As an important branch of computer vision,the target detection technology,has been a hot topic of research.As efficient recognition algorithms,deep learning and transfer learning,its characteristics include simple structure,strong robustness,less training parameters,etc.The image data are varied,complicated,and containing the variety of information,so it has become the focus and difficulty of the research,which is how to extract the useful information quickly and accurately by using relevant artificial intelligence technologies.In order to improve the accuracy and efficiency of a basic model of target detection which is called SSD,we adopt some mature image methods to process the relevant image data.It is combined the basic model with a lightweight deep neural network to construct a multi-scale convolutional neural network structure with feature pyramid.Firstly,this paper to collect and process the image data for the experiment needs.Secondly,to improve the basic model of object detection by training,verifying and testing the model with the collected image data,and make the model has a great robustness in multi-scene.Furthermore,the model is applied to a single pedestrian detection scene through the method of transfer learning,so that it has a good detection effect.The paper mainly focuses on the experiment and discussion of the following parts:?1?To reserve more target feature information and ensure the accuracy of model in the process of model detection,it amplified the feature map of the low-layer convolution layer,and then extracted the feature information of the high layer.While filtering the overlap target candidate areas,it set the threshold value to eliminate redundant target candidate areas based on the idea of non-maximum suppression,it made to reduce the number of negative samples and make the model's effect become stable gradually.In the target detection process of matching,to ensure the stability of the model,it conducted the positive and negative samples between the predicted region and the real region.?2?The activation function is extremely important in deep learning technology.In recent years,deep learning technology has made more and more achievements in various fields,and they can't be separated from the continuous development of activation function.Now,with the depth of the study,the existing activation functions,such as Sigmoid,Tanh,and ReLu,are exposed to more and more problems,for example,they are not robust to noise,and not favorable for neural network to learn sparse image data,even appear the overfitting because of “vanishing gradient” and “neuronal death”.The ReLu₆ is used to replaced the traditional ReLu function to prevent the phenomenon of overfitting.The experimental results show that ReLu₆ can efficaciously ease the phenomenon of overfitting,and make the model has a better robustness.?3?The basic convolutional neural network model structure has been improved.Aiming at the problem,such as the large number of parameters,the long operation time,the large consumption of computing resources,it combined with a lightweight deep neural network structure called MobileNet,to greatly reduce the number of parameters,and it has little effect on the decrease in model accuracy.Not only does it ensure the accuracy of the model,it also promote the speed and satisfies the real-time requirement of the model.In conclusion,this paper has carried out some research on deep learning in the field of target detection of computer vision.The proposed method has obtained certain experimental results,proved the effectivity of the proposed method.At the same time,it also demonstrates that the research results have a certain theoretical significance and the practical application value.