Pedestrian Information Collection And Processing System Based On Deep Learning

The video information collection and processing system refers to the process of capturing image information through a camera,and then analyzing or storing the captured image manually or by a machine.It is usually used for video surveillance or video abnormal behavior analysis.Among the collected video information,pedestrian information is a more important one.Using the facial information in the pedestrian information,you can quickly determine the identity of the pedestrian to ensure the safety of specific occasions.In traditional information collection and processing systems,most pedestrian information relies on manual screening and extraction,which is not only time-consuming but also laborious.When a pedestrian's face is covered,it is also difficult for humans to integrate information,which reduces the security and robustness of the entire system.And with the increase of video time,it brings the difficulty of storing large amounts of information.Therefore,in order to improve the existing collection methods,improve the accuracy and robustness of information collection,and simultaneously solve the massive information storage,this paper designs and implements a deep learning-based pedestrian information collection and processing system.The main tasks include:(1)According to the specific scene of pedestrian information collection,analyze and design the software architecture of pedestrian information collection and processing system based on deep learning,establish a four-layer model,presentation layer(human-computer interaction),business layer(image processing),persistence layer(Data storage),data layer(data type).According to the required configuration of the system,the hardware is selected in a targeted manner,and a reasonable hardware model architecture is designed.Complete the fisheye camera comparison,analyze the existing deep learning framework,and select the image processing model training method(2)For fast RCNN model,there is a problem of decreasing the map when video stream is input.On the basis of fast-rcnn,this paper designs and implements a non maxima suppression method of fusion time series,which includes dividing time series range,calculating reference frame score,time series re score,and high score removal.In order to reduce the score of strong detection frame(repeat frame),enhance the score of weak detection frame,complete the time sequence information fusion in NMS,and solve the problem of reducing the map in the detection process.(3)When the face is occluded,the integrity of pedestrian information is reduced.In this paper,a face symmetry algorithm based on gdbt is implemented to judge whether the face is occluded.A face completion algorithm based on wgan-gp is designed and implemented,and the local loss and global loss of the image are fused to improve the authenticity and similarity of the completed image and the original image in the process of image completion.Ensure the integrity of pedestrian information.(4)When storing a large amount of pedestrian information,the conventional storage method(centralized storage)is difficult to store.This paper analyzes the composition and storage requirements of stored data,compares common storage frameworks,selects Hadoop as the storage module of the information collection system,completes the architecture design of the information storage module,implements the Hadoop cluster,and stores pedestrian information on Hadoop,Display data storage results,and compare with common file transfer protocols to complete related storage performance tests.This completes the file storage.Finally,the performance of the system in this paper is tested through experiments Including the combination of different algorithms to collect pedestrian information integrity test,and the use of actual pedestrian data,information storage module stress test.The integrity test shows that in the Caltech test set,the algorithm combination in this paper is superior to other existing algorithm combinations,and the average recognition integrity is improved.The information storage test shows that the distributed storage system based on Hadoop has an average throughput of 77.8MB/s at the time of three nodes,which satisfies the demand of massive video storage.