A Spatiotemporal Multi-Granularity Perspective on Predicting Urban Sparse Traffic Accidents PROJECT TITLE : Foresee Urban Sparse Traffic Accidents: A Spatiotemporal Multi-Granularity Perspective ABSTRACT: Due to the rapid pace of urbanization, car accidents have evolved into a significant threat to both health and development. An accurate urban accident forecast enables higher-quality pre-allocation of police force and safe route planning for both traffic administrations and travelers, thereby reducing injuries and property damage to the greatest extent possible. Short-term accident forecasting methods that are available off-the-shelf typically operate on hourly levels and with a single step. These methods place their primary emphasis on modeling static region-wise correlations using existing neural networks. However, because of the ever-changing nature of road networks and the growth of urban areas, it can be difficult to improve the spatial and temporal granularity of forecasting. This is because of the rarity of accident records and the complexity of long-term dependencies on the outcomes of current decisions. We propose a unified framework called RiskSeq as a solution to these challenges so that we can anticipate sparse urban accidents on a finer granularity and with multiple steps from a spatiotemporal perspective. In particular, we design region-wise proximity measurements and temporal feature differential operations, and then embed them into a novel Differential Time-varying Graph Convolution Network to dynamically capture traffic variations. This allows us to better understand how traffic patterns change over time. In light of the hierarchical spatial dependencies and the obvious influences of context, a hierarchical sequence learning structure was developed by incorporating contextual factors into a step-wise decoder. This structure was created in consideration of the obvious context influences. In order to improve risk predictions that are founded on risk-gathering and risk-assigning networks, the multi-scale spatial risks are learned jointly. Extensive testing has shown that utilizing our RiskSeq can improve performance on two different datasets by anywhere from 5 to 15%. Did you like this research project? To get this research project Guidelines, Training and Code... Click Here facebook twitter google+ linkedin stumble pinterest Improved memory efficiency with fully dynamic kk-Center clustering FastDTW approximates the algorithm and is typically slower than the algorithm it approximates.