Parallel Feature Fusion Based On Choquet Integral And Its Application In Pedestrian Detection

In the construction of smart city,smart transportation is the most basic part.Intelligent transportation needs to build integrated traffic data analysis with decision support system,intelligent traffic management system,intelligent vehicle control system and vehicle network ecosystem.All the important systems mentioned above are inseparable from real-time pedestrian detection.However,human body has the characteristics of rigid and flexible objects,hence the appearance and posture of human body change greatly.It makes difficulties for scientists to achieve industrial-level speed and accuracy of pedestrian detection.Therefore,it is significant to improve the speed,accuracy and robustness of pedestrian detection.With regard to pedestrian detection,feature extraction and location based on classification are to core issues.This thesis mainly focuses on these two kernel aspects of pedestrian detection.The research strategy is as follows.Based on the theoretical and applied research of feature fusion using Choquet integral aggregation model,the key information contained in several commonly used pedestrian detection descriptors is accurately fused to explore the dynamic combination of feature fusion and pedestrian detection.A series of experimental analysis has been carried out.Combining prototype implementation with performance evaluation,this thesis systematically studies several key factors and problems of pedestrian detection based on Choquet integral.In feature extraction,a single feature descriptor cannot accurately describe pedestrian information.Thus,hybrid features become the mainstream of pedestrian detection.However,the current hybrid features are mainly implemented in serial fusion,which brings many difficulties and problems to pedestrian detection technology.On the one hand,serial fusion affects the speed of pedestrian detection and makes it hard to meet real-time requirements.On the other hand,serial fusion ignores the interaction between feature descriptors and influences the final accuracy.Therefore,a parallel feature fusion method based on Choquet integral is proposed in this thesis.In this thesis,a variety of low-level feature descriptors(mainly HOG and LBP)are studied.First,we focus on the establishment of Choquet integral model and the dimension matching of multiple low-level feature descriptors.Then,the parallel fusion flow-level feature descriptors are carried out by using Choquet integral aggregation model.In terms of location and classification,a pedestrian detection system is constructed based on the feature descriptor of parallel fusion of Choquet integral aggregation model and support vector machine(SVM)classifier.Finally,based on the benchmark pedestrian detection data set,the performance of existing research methods has been compared with the proposed method in this thesis.The experimental results show that our method not only improves the detection rate of pedestrian detection but also reduces the extraction time of features.It makes the pedestrian detection easier to meet the real-time requirements,and has a good effect in solving the problem of pedestrian occlusion.