ARIS: A Real Time Edge Computed Accident Risk Inference System

To deploy an intelligent transport system in urban environment, an effective and real-time accident risk prediction method is required that can help maintain road safety, provide adequate level of medical assistance and transport in case of an emergency. Reducing traffic accidents is an important problem for increasing public safety, so accident analysis and prediction have been a subject of extensive research in recent time. Even if a traffic hazard occurs, a readily deployable structure with an accurate prediction of accident can contribute to better management of rescue resources. But the significant shortcomings of current studies are the use of small-scale datasets with minimal scope, being based on extensive data sets, and not being applicable for real-time purposes. To overcome these challenges, we propose ARIS: a system for real-time traffic accident prediction built on a traffic accident dataset named ‘US-Accidents’ which covers 49 states of United States, collected from February 2016 to June 2020. Our approach is based on a deep neural network model that utilizes a variety of data characteristics, such as time-sensitive weather data, textual information, and discerning factors. We have tested ARIS against multiple baselines through a comprehensive series of experiments across several major cities of USA, and we have noticed significant improvement during inference especially in detecting accident classes. Additionally, to make our model edge-implementable we have compressed our model using a joint technique of magnitude-based weight pruning and model quantization. We have also demonstrated the inference results along with power consumption profiling after deploying the model on a resource constrained environment that consists of Intel Neural Compute Stick 2 (NCS2) with Raspberry Pi 4B (RPi4). Our investigation and observations indicate major improvements to predict unusual traffic accident event even after model compression and deployment. We have managed to reduce the model size and inference time by ≈ 6x, and ≈ 70 % respectively with insignificant drop in performance. Furthermore, to better understand the importance of each individual type of variables used in our analysis, we have showcased a comprehensive ablation study.