文章基本信息

标题：Reward based Action Learning from Vehicle Mounted Camera Images for Real-Time Simulation and Rendering of Indian Traffic
本地全文：下载
作者：Dr. G. Murugaboopathi ; G.Sankar
期刊名称：International Journal of Computer Science & Technology
印刷版ISSN：2229-4333
电子版ISSN：0976-8491
出版年度：2012
卷号：3
期号：4
页码：464-466
语种：English
出版社：Ayushmaan Technologies
摘要：A traffic simulation model with detailed simulation and realistic rendering serves the purpose of global traffic prediction as well as individual driver behavior analysis. Though there has been research in traffic simulation for years, the focus has been mainly on centralized approaches. The main drawback with the centralized approaches is that they are quite intractable, unstable and unrealistic. Recently many decentralized models are developed to reproduce global parameters like traffic flow rate from collective behavior of intelligent agents. But decentralized approaches which simulate local parameters like individual speed preferences do not consider the lane-less chaotic traffic existing in most Indian roads. They do not provide sufficient details for realistic rendering. As of now, there are no detailed simulation models available for realistic rendering that captures the essential features of Indian traffic scenario. The proposed work tries to avoid the above mentioned problems and aims to develop a decentralized system for detailed simulation and realistic rendering of Indian traffic. Some of the nationally significant applications of the proposed tool are urban planning, training novice drivers and serving as a test bed for new traffic implementations apart from games and entertainment. From fig. 1, it can be seen that research in traffic monitoring and simulation was started a century ago. Over the years, advanced techniques such as loop detectors and sensors have refined the traffic simulation models. Recently with the availability of devices and tools to capture and process real-time videos, there are new inputs to the learning algorithms used in traffic simulation. Loop detectors and road-side cameras are known to provide global traffic information such as vehicle type and density in a road, traffic jam and queue length at traffic junctions. But they are not sufficient to develop detailed simulation models which require realistic rendering. Though vehicle velocity may be measured using vehicle mounted devices, that alone is not sufficient for the task. We need to learn the driver actions with respect to the acceleration and break controls to simulate the realistic path of the vehicle on the road. This emphasizes the need to use vehicle mounted cameras for capturing vehicle traffic on a road.
关键词：Artificial Intelligence;Reinforcement Learning;Traffic Simulation; Realistic Rendering;Object Tracking