Road Accident Anticipation Using Dashcam Video with Deep Learning Models

Ooi, Presley Yit Hoong (2025) Road Accident Anticipation Using Dashcam Video with Deep Learning Models. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Presley - FULL TEXT.pdf Restricted to Repository staff only Download (2MB)

Abstract

Traffic accident anticipation (TAA) using dashcam videos is vital for autonomous vehicles (AVs) and advanced driver-assistance systems (ADAS) to pre-emptively address road safety challenges, given the significant global toll of traffic accidents caused largely by human error. However, existing TAA models often suffer from low precision or insufficient Time-to-Accident (TTA), limiting their effectiveness in dynamic, high-speed, or complex driving environments where longer anticipation windows are critical for intervention. To tackle this, this research proposes a novel deep learning framework that integrates a Dynamic Spatial Attention (DSA) network with a Dual Recurrent Neural Network (Dual-RNN) Averaging Model combining Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) to enhance early TAA with improved average precision (AP). The DSA network dynamically prioritizes critical spatial regions in dashcam frames, while the Dual-RNN model captures both long-term and short-term temporal dependencies with an element-wise averaging model to enhance prediction robustness. This approach introduces a pioneering integration of spatial attention and dual RNNs, significantly extending time-to-accident (TTA) and improving prediction accuracy. Compared to other models, this model achieves state-of-the-art performance, demonstrating balanced improvements in both AP and TTA in urban area datasets such as Dashcam Accident Dataset (DAD) with an AP of 75.73% and a TTA of 2.224s while maintaining substantial results for Car Crash Dataset (CCD) with an AP of 99.42% and a TTA of 4.326s. By advancing anticipation capabilities, this framework provides a scalable, real-time solution for accident prevention in AVs and ADAS, contributing meaningfully to global road safety efforts.