Multi Feature Fusion Based Bus Passenger Re-Identification Research And Application

With the development of artificial intelligence technology,pedestrian reidentification algorithms based on multi-camera and multi-time-point video information in public environments have been widely researched and applied in the field of intelligent security.By using multi-source and time-based feature information,these algorithms can recognize people’s identities and associate them with spatiotemporal trajectories.Utilizing multiple cameras on public transportation vehicles,such as buses,for passenger re-identification,not only allows for precise identification of passengers’ identities through multiple sources of information,but also enables accurate positioning of passengers’ boarding and alighting locations.This approach can establish precise passenger flow models,promoting the development of intelligent transportation,and enable the identification and tracking of specific groups,enhancing public safety.However,due to the long working hours,high passenger density,and issues such as insufficient lighting,lighting changes,occlusions,and clothing changes,passenger re-identification tasks in the public transportation environment are highly challenging.To address the aforementioned issues,this thesis employs the three-dimensional features of the human body and other modal information to compensate for the limitations of single visible light,and conducts a study on bus passenger reidentification based on multi-feature fusion.Firstly,in order to reduce the storage cost of depth information,speed up the processing speed and throughput on CPU-constrained devices,a parallel deep video compression algorithm named Zdepth-Parallel has been designed.Compared to the existing algorithm Zdepth,this algorithm improves the speed by 260% on Jetson Xavier NX,thereby ensuring the real-time processing of depth video.Secondly,in order to overcome the limitations of single modality in the bus scene,infrared modality and depth modality were introduced based on visible light modality,and relevant datasets were constructed.A Triple Cross Fusion model that fuses the features from all three modalities using feature attention was designed.By introducing attention mechanisms,the network is able to focus on modalities with rich features,and the mAP metric is improved by 3.34% compared to existing visible light methods.This enhances the accuracy of re-identification tasks in bus scenes.Once again,to enhance the 3D features of the human body without the depth information,a re-identification model based on 3D reconstruction has been established.The model uses human body 3D reconstruction technology to extract normal information from the front and back of the body,and combines the extracted normal features with visible light through feature fusion.In the Market1501 dataset,the network with this branch added improved the mAP indicator by 1.36% compared to the existing ConvNeXt network.Finally,to make the re-identification model proposed in this thesis more applicable to real-world scenarios,multiple acceleration methods were introduced to optimize the model.At the same time,a joint process of object detection and pedestrian reidentification was designed,as well as a graphical interface to ensure real-time and readability of the application.The results showed that after the model was accelerated,its inference speed was improved by 55.1% compared to before acceleration.