Pedestrian Tracking Method Based On Feature Fusion

Pedestrian tracking is a hotspot in the field of computer vision,which aims to track specific pedestrian targets in video sequences.The current mainstream algorithms of pedestrian tracking mainly use the appearance characteristics for positioning,when encountering object occlusion,blurred light,complex changes,etc.,the accuracy will be reduced and the pedestrian target will be lost and difficult to be retrieved,which is hard to meet the actual needs.In order to solve the problems,this thesis focuses on the research on the pedestrian tracking based on feature fusion,and the main contents and contributions are as follows:A cross-modal target segmentation method based on attention mechanism is proposed to eliminate the influence of background chaos on target positioning.The method is based on DMN,firstly,the CBAM attention mechanism is introduced in the visual feature extraction backbone network,which extracts and optimizes the features in the dimensions of channel and space respectively to suppress the irrelevant noise;secondly,the joint regularization of FRN and BN is introduced to suppress the impact of batch size on network performance;next,the optimized visual features are fused with the language features and continuously up-sampled to obtain the target segmentation results.Finally,MIo U is used as an evaluation metric for simulation experiments,showed that the proposed model improved by 1.85% and 0.52% on the Refer It and GRef datasets respectively.A pedestrian tracking method based on the fusion of appearance and action characteristics is proposed and the Da Siamese-RPN is used as the baseline model,which aims to eliminate the negative interference caused by the appearance deformation when the target is moving.Firstly,the appearance and action characteristics fusion module is introduced in the sub-network to fuse the spatiotemporal information and action characteristics of the target;secondly,the global and local differences between the real label and the predicted template frame are calculated,which are weighted fused to construct the loss function,aiming to update the template frame online;finally,the classification and regression scores of the target are calculated to get the position of the target.Simulation results showed that the tracking success rate and precision rate have improved by 0.6% and 1.7% respectively on the La SOT dataset,and have improved by 0.9%and 0.8% respectively on the OTB100 pedestrian dataset.A pedestrian tracking method based on contextual information fusion is proposed and the SNLT is used as the baseline model,aiming to eliminate the interruption due to the complexity of the external environment.Firstly,the static context information of the image is fully obtained by convolution calculations,therefore,the attention value is calculated and the dynamic context information is updated;then,the dynamic and the static context information are fused to update the visual features;finally,calculate the classification and regression scores of the extracted features to get the tracking results.Simulation results showed that the success rate and the precision rate have improved by 2.1% and 0.8% separately on the pedestrian dataset OTB-99-LANG,and have improved by 1.1% and 1.2% separately on the public dataset La SOT.In terms of software implementation,the Pytorch deep learning framework is adopted,the core algorithm is programmed through the Python language,the interface development is carried out by Matlab R2016 b.The pedestrian tracking software has the functions of pedestrian positioning and pedestrian tracking,which is validated on public datasets and videos of real scenes.