Multi-view Object Detection Based On Balanced Graph Matching

In recent years, multi-view target detection has gained widespreadattention. However, there are many problems in the process of targetdetection as follows:(1) Multi-view motion state to estimate problem hasa very high computational complexity, and its state space is calleddimension disaster, which increase exponentially with the viewing angleand the target number.(2) Shelter increases the difficulties of themulti-view target detection and tracking.(3) The majority of the existingframework need to pre-calibrate the multi-view camera system, and therequirements for the accuracy of the calibration result is quite high.Based on extensive analysis of relevant research at home and abroad,this study deep analyzed the questions above in multi-view targetdetection, furthermore carried on algorithm design and implementationaccording to proposed theories. In this study, the existing multi-viewalgorithms was summarized, multi-view detection probabilisticframework basing on maximum a posteriori was found, and the variationapproximation was used to solve the problem of high-dimensional statespace introduced by the joint posterior probability and local detectionblock. In order to solve problem about the description of thecorresponding probability in different perspectives nodes to inter-viewgeometric constraints, this study constructed the inter-perspective imagematching function to describe the corresponding position potentialfunction, improved the multi-view target detection framework, usedmatching function to describe the location potential function incharacteristic transfer link with different perspectives. In the matching process, Compatibility matrix greatly improved the robustness of imageregistration by using the bistochastic normalization optimization. In localdetection, the use of algorithm to optimize the original framework, suchas TLD, could improve the detection accuracy. The experimental resultsindicate that the multi-view object detection algorithm basing on balancedgraph matching can carry on the multi-view target detection with higherprecision and stronger robustness.