Vehicle Re-reidentification Based On Multiscale Information Fusion And Multi-angle Visual Surface Matching

With the rapid development of information technology,the generation of traffic monitoring data is getting faster and faster,and its application is becoming more and more popular.under this background,people put forward the concept of smart city.Smart City aims to process a large amount of data generated in the city through high-performance calculators,reduce management costs and liberate the productivity of managers and decision makers.Vehicle rerecognition is one of the important analysis technologies in intelligent city monitoring and safety management.Vehicle re-identification(Re ID)algorithm needs to make great use of the subtle appearance information in the image data,so whether it can fully represent the highresolution information in the vehicle picture becomes the key to solve the problem.Aiming at the vehicle re-recognition data set collected from the open scene,this thesis proposes a series of algorithms to solve the current vehicle re-recognition problem.First of all,this thesis observes the characteristics of the data,and analyzes the reasons for the low performance of the general algorithm and the current difficulties in vehicle re-recognition: the algorithm is required to be robust enough.the correct appearance representation must be constructed under different viewing angles,resolutions,occlusion and lighting conditions.However,the appearance of the same vehicle is very different under different conditions,but the appearance of different two vehicles is very similar.In view of these difficulties,this chapter proposes two different vehicle re-recognition algorithms to solve the two general problems of “scale unbiased” and “visual angle unbiased” respectively.For the "scale unbiased" problem,this thesis proposes a vehicle re-recognition algorithm based on multi-scale information fusion based on self-attention mechanism.In this algorithm,a multi-head self-attention mechanism is added to the specific network layer of the backbone network.this mechanism constructs the energy activation graph through weight calculation,and constructs the relationship between different spatial locations.And use the two characteristics of the global maximum pooling layer to extract the edge and the global average pooling layer to extract the material,the branches in different stages carry on the combined pooling operation to the features of different granularity,and stack the output of multiple branches.finally,the full connection layer is used to output the final unified representation vector.The postprocessing uses the reordering algorithm to optimize the results.For the problem of unbiased viewing angle,a vehicle re-recognition algorithm based on image segmentation(UNet)is proposed in this thesis.In this algorithm,each visual surface of the vehicle is segmented in the data pre-processing part,and the global feature representation of the vehicle is obtained by processing the vehicle image through the backbone network and multi-head attention mechanism.After that,the visual surface segmentation mask obtained by pre-processing is used to ROI-align the global features,and the corresponding feature representation of each visual surface is obtained.finally,the similarity between the same visual surfaces is calculated.This operation can not only extract the vehicle body from the complex background,but also accurately match the different faces of the vehicle.The experimental results show that the performance of the algorithm can be greatly improved by finding important regions through attention module,constructing feature vectors with different receptive field granularity and calculating similarity between visual surfaces.Finally,this thesis makes a summary and prospect of the proposed algorithm and the future development of the field.