In-Memory Deep Learning Computations on GPUs for Prediction of Road Traffic Incidents Using Big Data Fusion

A staggering 1.25 million people die and up to 50 million people suffer injuries annually due to road traffic crashes around the world, causing great socio-economic and environmental damages. Road collisions are a major cause of road congestion. The cost of congestion to the US economy, alone, exceeded 305 billion USD in 2017. Smart infrastructure developments have accelerated the pace of technological advancements and the penetration of these technologies to all spheres of everyday life including transportation. The use of GPS devices to collect data, image processing and artificial intelligence (AI) for traffic analysis, and autonomous driving are but a few examples. This paper brings together transport big data, deep learning, in-memory computing, and GPU computing to predict traffic incidents on the road. Three different kinds of datasets—road traffic, vehicle detector station (VDS), and incident data—are combined together to predict road traffic incidents. The data is acquired from the California Department of Transportation (Caltrans) Performance Measurement System (PeMS). We have analyzed over 10 years of road traffic data. This work-in-progress paper reports incident prediction results using 3 months’ data, September to November 2017. The data fusion methodology is explained in detail along with the algorithms. The results for various configurations of deep convolution neural networks are given. Conclusions are drawn from the current status of the results and ideas for future improvements are given.