Font Size: a A A

Research On Pedestrian Recognition Technology Based On Deep Learning And Representation Design

Posted on:2021-05-23Degree:MasterType:Thesis
Country:ChinaCandidate:Y HanFull Text:PDF
GTID:2518306554465064Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
Pedestrian re-identification is applied in the field of public safety.The purpose is to identify and match the same person across scenes.A set of efficient and accurate pedestrian re-identification algorithms will help relevant personnel to better query the target person through the existing video surveillance system.Pedestrian re-identification mainly solves the problem of pedestrian retrieval and positioning under different regional backgrounds.The research difficulties include:pedestrians cannot be well aligned during comparison and recognition,there are many noisy areas in the image,and the data sets that can actually be obtained are video segments extracted from the camera,which consists of a series of pedestrian frames of continuous images Composition,there is a correlation between each picture,and the number is huge,which consumes a lot of resources when processing these data sets.In response to these problems,this article mainly solves from the following aspects.Firstly,a pedestrian re-recognition method based on area alignment is proposed.The pedestrians are located by head detection,and the image is segmented.Deep learning is used to extract features from the image.The pedestrian recognition rate is calculated by Softmax.Solve the problem of low accuracy of pedestrian reidentification due to misalignment and image noise.(1)Design a head detection model that can locate the area in the image where the pedestrian is based on the head.(2)A feature extraction network is constructed based on the resnet-50 network,which is used to obtain the input pedestrian overall feature vector.(3)Design a segmentation network to segment the extracted overall features into local features and further convolution for training.(4)Combine the overall features and local features as pedestrian identity features and use the softmax loss function for training.Experiments show that the method uses the Res Net framework to test on Market1501 pedestrian re-recognition data set,and the accuracy rate can reach Rank1 to 93.8%,which is better than the currently performing pedestrian re-recognition based on gesture segmentation.Then in order to deal with the huge data set based on video pedestrian rerecognition research,and because the data set is composed of a series of consecutive frames of pedestrian images,there is a correlation between each picture.In feature time,there is an average processing of consecutive frames,which cannot deal with the problem of temporal continuity and spatial alignment at the same time.A pedestrian reidentification time modeling method based on non-local attention mechanism is proposed.(1)A video pedestrian recognition network is designed,including image feature extractor,temporal modeling method,and loss function for training.(2)The video data set is clipped into fragments with a fixed number of frames,and the fragments of pedestrians are randomly extracted as the input of the network,which can maximize the use of all data.(3)Use non-local attention mechanism to model,expand the receptive field of the model on the image,so that the model can notice more information.(4)The model takes continuous frames as input,outputs the segment feature of pedestrians in the previous layer of the classification layer,and uses the softmax function and the triplet function to jointly train the network.Experiments show that the non-local attention mechanism can calculate the interaction dependence between arbitrary positions on the image,and is not limited to adjacent points,so it can maintain more information when the pedestrian starts to drift over time to solve the problem.The problem of spatial alignment during feature fusion.In this method,under the video pedestrian data set MARS,the R-1 of this paper can reach 84.8%,which is 5% higher than the average time pooling method and 1.5% higher than the previous attention-weighted average pooling method..The result proves that the method in this paper can effectively solve the problems of temporal continuity and spatial alignment in video pedestrian recognition.In summary,this paper has conducted research and exploration on key technologies for pedestrian re-identification from different data sets,and these technologies can be applied in the field of video surveillance.
Keywords/Search Tags:pedestrian recognition, head detection, pedestrian alignment, temporal modeling, attention mechanism
PDF Full Text Request
Related items