Video based Traffic Forecasting using Convolution Neural Network Model and Transfer Learning Techniques

Abstract

The ideas, algorithms and models developed for application in one particular domain can be applied for solving similar issues in a different domain using the modern concept termed as transfer learning. The connection between spatiotemporal forecasting of traffic and video prediction is identified in this paper. With the developments in technology, traffic signals are replaced with smart systems and video streaming for analysis and maintenance of the traffic all over the city. Processing of these video streams requires lot of effort due to the amount of data that is generated. This paper proposed a simplified technique for processing such voluminous data. The large data set of real-world traffic is used for prediction and forecasting the urban traffic. A combination of predefined kernels are used for spatial filtering and several such transferred techniques in combination will convolutional artificial neural networks that use spectral graphs and time series models. Spatially regularized vector autoregression models and non‐spatial time series models are the baseline traffic forecasting models that are compared for forecasting the performance. In terms of training efforts, development as well as forecasting accuracy, the efficiency of urban traffic forecasting is high on implementation of video prediction algorithms and models. Further, the potential research directions are presented along the obstacles and problems in transferring schemes.