ENHANCED DEEP LEARNING-BASED REAR COLLISION AVOIDANCE SYSTEM USING A HYBRID CNN–RNN ARCHITECTURE AND MULTI-SENSOR FUSION

Abstract

Rear-end collisions remain a major contributor to road traffic accidents, accounting for nearly 28% of all road accidents in Nigeria in 2022, with existing collision avoidance systems limited by rigid, rule-based logic and reliance on costly or single-modality sensing. This work enhances a Raspberry Pi 4B-based Rear Collision Avoidance System (RCAS) that originally relied solely on YOLOv8-based object detection and achieved an average precision of 80.9%. The enhancement introduces a hybrid three-dimensional convolutional neural network (3D-CNN) and Long Short-Term Memory (LSTM) architecture for temporal awareness, retains YOLOv8-nano for real-time object detection, and incorporates a confidence-based ultrasonic–camera sensor fusion scheme to improve robustness. The system was implemented and evaluated on a Raspberry Pi 4B (4 GB RAM) platform using a synthetic road-scene dataset of 500 annotated images. Experimental results show that the hybrid CNN–RNN module achieved 88% temporal prediction accuracy, YOLOv8-nano attained 85% detection precision and 82% recall at 25 frames per second, and sensor fusion raised the overall detection rate to 92% while reducing the false alarm rate from 15% (camera-only) to 4%. The complete system attained an overall performance score of 0.874 at an estimated power consumption of 5.5 W, confirming its suitability for low-cost, real-time embedded deployment. These results demonstrate that combining temporal deep learning with multi-sensor fusion measurably improves collision detection accuracy, robustness, and responsiveness over the existing single-modality baseline.

Keywords: Rear collision avoidance, convolutional neural network, recurrent neural network, LSTM, YOLOv8, sensor fusion, ultrasonic sensor, Raspberry Pi, embedded systems, advanced driver assistance systems.