详情

原文传递 Super synchronization for fused video and time-series neural network training

题名：	Super synchronization for fused video and time-series neural network training
作者：	Gillmann, R.;Pepin, J.;Elliott, C.J.;
关键词：	99 MATHEMATICS, COMPUTERS, INFORMATION SCIENCE, MANAGEMENT, LAW, MISCELLANEOUS; NEURAL NETWORKS; SYNCHRONIZATION; TRAFFIC CONTROL; TIME-SERIES ANALYSIS
摘要：	A key element in establishing neural networks for traffic monitoring is the ground truth data set that verifies the sensor data. The sensors we use have time series data gathered from loop and piezo sensors embedded in the highway. These signals are analyzed and parsed into vehicle events. Features are extracted from these sensors and combine to form the vehicle vectors. The vehicle vectors are combined together with the video data in a data fusion process thereby providing the neural network with its training set. We examine two studies, one by Georgia Tech Research Institute (GTRI) and another by Los Alamos National Laboratory (LANL) that use video information and have had difficulties in establishing the fusion process. That is to say, the correspondence between the video events recorded as the ground truth data and the sensor events has been uncertain. We show that these uncertainties can be removed by establishing a more precise and accurate time measurement for the video events. The principal that the video time information is inherently precise to better than a frame (1/30 s) and that by tracing the factors causing imprecision in the timing of events, we can achieve precisions required for unique vehicle identification we call super synchronization. In the Georgia data study there was an imprecision on the order of 3 seconds and in the LANL study an imprecision of early a second. In both cases, the imprecision had led to lack of proper identification of sensor events. In the case of the Georgia 120 study sensors were placed at various distances downstream, up to 250 meters, from the ground truth camera. The original analysis assumed that there was a fixed time offset corresponding to the downstream location. For this case we show that when we restrict the analysis to passenger cars and take into account the speed of the car we can achieve a precision of approximately 0.3 s. This value is an order of magnitude less than the previous procedure.
总页数：	12p
报告类型：	科技报告