Dynamic Scene Understanding from Dashcam Videos using YOLO Detection, Monocular Depth Estimation, and Attention-Based Autoencoders

Road safety remains a critical concern, with the detection and prediction of accidents from dashcam footage becoming increasingly vital. In this work, we present an end-to-end system that integrates object detection, monocular depth estimation, and attention-based autoencoders to understand dynamic road scenes. The proposed pipeline employs YOLOv8s for object detection, ZoeDepth for depth estimation, and a Transformer-inspired attention autoencoder to model inter-object relations in an unsupervised manner. Rather than directly classifying accident events, our model highlights accident-prone areas through learned attention maps. Results demonstrate strong reconstruction performance, achieving 98% accuracy, and effective visualization of risk concentrations. This fusion of detection, depth, and attention-driven reasoning offers a new perspective on proactive accident scene analysis.