Pedestrian Detection And Trajectory Analysis Based On Text Guidance

Video-based pedestrian trajectory analysis is an important part of public security work,and pedestrian detection and person re-identification are the key technologies.At present,the mainstream method of pedestrian detection is to use the appearance characteristics of pedestrians for positioning,which does not make full use of the text description information for specific pedestrians.Affected by the change of real scene,the pedestrian feature discrimination extracted by the model is not strong,and the robustness of the model to image corruption is not high.In view of the above problems,the main research content and contributions of this paper are as follows:An improved pedestrian detection method of Retina Net is proposed.Firstly,a text feature extraction module is constructed to extract text features from the text description of target pedestrian,and generate corresponding dynamic filters.Secondly,a cross-modal feature fusion module is constructed,which fuses the extracted visual features and text features to generate a feature response map for a specific target.Then,the upsampling module is constructed to combine the original visual features and the feature response map of a specific target to generate a text-guided target pedestrian proposal.Finally,distinguish between the pedestrian proposal generated by the Retina Net pedestrian detection network and the target pedestrian proposal outputted by the upsampling module to obtain the optimal result and achieve accurate positioning of specific pedestrian.Extract pedestrian images from the MSCOCO dataset for training and testing,and output visualization results.The experimental results show that the proposed method realizes the joint supervision of text and image,focusing on specific pedestrian target.A person re-identification method based on attention block and conditional convolution is proposed.Based on the CTL network model,firstly,the CBAM attention module is embedded in the backbone to weighted and strengthen the key information on the input image space and channel.Secondly,conditional convolution is introduced into the backbone,and the convolution kernel parameters are dynamically adjusted,so that the model can improve the model capacity while maintaining efficient inference.Finally,the weighted features are used to calculate the average intraclass centroid of pedestrians,which effectively reduces the search space of pedestrian targets to be queried and further improves the retrieval efficiency.In the evaluation on Market1501,MSMT17 and Duke MTMC-Re ID datasets,the evaluation indexes Rank1 increased by 1.1%,2.4% and 1.3%,and m AP increased by 0.5%,2.3% and 1.3%,respectively.Experimental results show that the improved method can effectively excavate pedestrian features with strong characterization ability.A person re-identification method based on joint normalization and deep connection attention strategy is proposed.Based on the CIL network model,a joint normalization strategy is first introduced in the feature extraction network to reduce the feature differences caused by image corruption.Then,the SE channel attention module based on deep connection strategy is embedded in the feature extraction network to enhance the feature extraction capability of the model.The improved model was evaluated by three image corruption methods on the Market1501,MSMT17 and CUHK-03 datasets,and the m INP increased by 2.66%,6.67% and4.14% on the Market1501 dataset,0.34%,1.21% and 0.32% on the MSMT17 dataset,and 1.8%,3.3% and 5.31% on the CUHK-03 dataset.Experimental results show that the improved model can further improve the robustness of image corruption.In terms of the implementation and verification of the method in this paper,the Pytorch deep learning framework is adopted,the core algorithm is implemented through the Python language,and the interface development is carried out by using Py Side2,and the simulation program based on method presented in this article is programmed,with functions such as text-guided pedestrian detection,person re-identification,corruption image person re-identification,pedestrian trajectory generation,etc.