A comparative performance analysis of accident anticipation with deep learning extractors

Abstract

Accident anticipation has become a major focus to avert accidents or to minimize their impacts. Over the years, several network systems are being developed and applied in self-driving technology. Despite the fact that advancement in the autonomous industry is fast-growing, major efficiency is required in the network systems that are gradually emerging. Recent research has proposed a novel end-to-end dynamic spatial-temporal attention network (DSTA) by combining a Gated Recur- rent Unit (GRU) with spatial-temporal attention learning network, to identify an accident video in 4.87 seconds before the occurrence of the accident with 99.6% ac- curacy when tested on the Car Crash Dataset (CCD). However, DSTA has not been able to provide efficient results on the Dashcam Accident Dataset (DAD) dataset. Moreover, the GRU model integrated in the DSTA network has a weak information processing capability and low update efficiency amid several hidden layers. The decision-making process of the accident anticipation network may be understood using the high quality saliency maps produced by the Grad-CAM and XGradCAM approaches. In this paper, we evaluate that using Wide ResNet network enhances the performance mechanism of feature extraction to increase accident anticipation precision. This change improves the capacity to process information and the learning efficacy. In addition, we suggest employing a Gated Recurrent Unit (GRU) network which will serve as a prominent feature to train the model to recognize data’s sequential properties and apply patterns to forecast the following likely event. Hence, we plan to incorporate Wide ResNet50, a system for extracting features which will identify the vehicles at risk by using wider residual blocks. These neural networks generate labels for identifying hazardous conditions in driving environments in order to anticipate accidents.