Crashes in traffic are a kind of geographical event, in which their spatial structure is generally linked to the regional pattern of human activity, sociodemographics and the road network. The sudden development in the huge amount of data collection, transformation and storage technology provides new techniques that can be efficiently used to enhance the prediction of traffic crashes. Therefore, it is significant to solve the issues that exist in the traditional crash prediction model. Hence, a novel deep learning-based crash prediction framework is implemented with big data. Initially, essential big data for the validation is collected from the benchmark resources. From the collected big data, essential features are extracted using a developed Spatial and Temporal Attention-based Autoencoder (STA-AE). Further, the extracted features are provided to the crash prediction phase. In the developed framework, crash prediction is executed through Adaptive Residual Bidirectional Long Short-Term Memory (ARes-BiLSTM). Moreover, the developed crash prediction model’s parameters are tuned using a Fitness-based Constant Circle Search Algorithm (FitCCSA). The developed ARes-BiLSTM offers crash-predicted outcomes. Further, various experiments are executed in the developed framework to verify its effectiveness over the classical techniques.
Dharsini et al. (Fri,) studied this question.