Research On Pedestrian Re-identification For Vehicle-road Collaborative Perception

With the rapid development of artificial intelligence technology,how to improve the safe driving ability of intelligent driving vehicles is a problem that researchers are committed to solving.However,the perception ability of a single intelligent driving vehicle is limited,and relying solely on a single vehicle for pedestrian detection to achieve pedestrian collision avoidance has a bottleneck in system reliability.The joint detection of pedestrians by the Collaborative Vehicle Infrastructure Systems(CVIS)can effectively solve the problem of visual limitations in the detection process,thereby improving the pedestrian collision warning capability of the driving assistance system.In order to solve the problem of vehicle-road collaborative perception of pedestrians,the feature extraction algorithm and feature retrieval algorithm of pedestrian re-identification(Re ID)on CVIS are mainly studied,and related algorithm optimization and experimental verification are carried out.In order to improve the accuracy and real-time requirements of the Re ID results for CVIS,feature extraction and dimensionality reduction algorithms are designed to extract effective and low-dimensional pedestrian image features for Re ID.The multi-task loss function is used to constrain the training of the multi-scale granularity network(MSGN),and the obtained model can extract the strong discriminative features of the pedestrian images and the accuracy rate of Rank-1 reaches 94.8%.The feature dimensionality reduction algorithm based on local sensitive hash compresses the high-dimensional floating-point deep features of the pedestrian image extracted by the MSGN into low-dimensional binary features,which can reduce the feature volume by 28 times under the situation that Rank-1 is only reduced by 0.43%,and achieve a significant reduction in the communication overhead caused by information interaction at both ends of the vehicles and roads.Moreover,a vehicle-road cooperative pedestrian Re ID system is constructed,which can realize the adaptive adjustment of the image feature dimension extracted according to the change of the wireless communication network channel state,making the Re ID result more suitable for the CVIS scene where communication network resources change with vehicle density in real time.In the aspect of feature retrieval,in order to solve the problem of low efficiency in the retrieval methods that use query images and all candidate images one-to-one matching in complex CVIS scenes,a retrieval optimization algorithm based on Hamming distance metric and spatio-temporal clustering is proposed.On the basis of low-dimensional binary image features,it is proposed to use Hamming distance to replace Euclidean distance for feature similarity measurement,which reduces the computational complexity of feature matching and increases the retrieval speed by 23 times.A clustering algorithm that combines spatio-temporal constraints which can optimize the feature retrieval method of Re ID is proposed.By designing a reasonable cluster center selection method and using spatio-temporal features to compensate for the limitations of visual features,the extracted candidate pedestrian image features are effectively classified,so that the images of the same pedestrian are divided into the same cluster.Executing pedestrian image retrieval tasks in the corresponding clusters reduces the number of times of similarity matching of pedestrian image features by narrowing the retrieval range,and increases the retrieval speed by 2 times while ensuring the accuracy of Re ID,which further improves the Re ID effectiveness.The experimental results prove that the proposed Re ID algorithm can be applied to the CVIS with real-time and accuracy demand,and provides a good theoretical and experimental basis for realizing the vehicle-road collaborative perception of pedestrians.