| Summary: | 碩士 === 國立交通大學 === 網路工程研究所 === 107 === Traffic conditions are closely related to vehicle speeds. When traffic is congested, the vehicle speeds will be severely affected, and vice versa. Therefore, it is very helpful for drivers to know the instantaneous speeds of the traffic ahead. Thanks to new technological advances, Traffic Information Systems (ITS) provide more and more traffic information. For example, CCTV videos can be processed to provide users with instantaneous vehicle speeds. This paper uses the CCTV videos provided by the Taiwan Freeway Bureau to estimate the traffic speeds in the Hsuehshan Tunnel, in order to provide more real-time traffic speed information.
For our dataset, we use CCTV videos as inputs, VD measurements of 1-minute mean speeds as the true values, and MAE to evaluate the errors. We use a combined 3DCNN and RNN architecture to learn spatiotemporal features. Finally, a fully-connected layer is used to generate the traffic speed. One of our investigations is to determine the optimal time-span that the 3DCNN should examine. We use four different depths of 3DCNN; the deeper the 3DCNN, the shorter time sequence the RNN learns; the shallower the 3DCNN, the longer time sequence the RNN learns. Our experimental results show that a moderate depth of 3DCNN followed by three-layer Gated Recurrent Units (GRUs) generates the best results. Compared with the existing architectures such as LRCN and single 3DCNN, the MAE of our method is reduced by 11% on average. We have tested on videos of 8 CCTVs, our prediction MAE can be as small as 3 km/hr., which is very accurate for practical traffic information applications.
There are two limitations in our research. First of all, we need to build two models to train CCTV data of different frame numbers. Secondly, we only estimate the CCTVs in the Hsuehshan Tunnel. In the future, we hope to overcome these two limitations and make the application more extensive.
|
|---|
