On feature selection in network flow based traffic sign tracking models

Abstract We use the network flow based tracking model to process videos recorded by taxicabs to collect information, such as the locations and types, of traffic signs. One of the key questions in this type of tracking models is the feature selection problem, which determines the features and their associated weights to construct the costs on the edges of the underlying network. However, in existing studies, the features and their weights are set mostly by intuition and there are no studies that systematically investigate this problem. We investigate the feature selection problem by considering a wide range of features related to traffic sign candidates, including its position, size, frame index, color histogram, speed up robust feature (SURF) descriptors, and template matching similarities. We propose to use the multinomial logit model to identify features that have significant influence on the edge costs and to estimate their corresponding weights. We test the performance of our selected features using real videos recorded by taxicabs in the city of Beijing. We find that the position and frame index of the traffic sign candidates are the top two features that influence the tracking results. We also find that image similarity measures are helpful features, but their contribution is not as large as the previous two features.