Highway Lane-Change Prediction and Recommendation on an Embedded Edge Platform

Abstract

Traffic accidents on Indonesian highways are largely driven by unpredictable lane-change maneuvers. While Advanced Driver Assistance Systems (ADAS) offer preventive solutions, their high computational requirements often restrict deployment to premium vehicles. This study focuses on the deployment and optimization of a real-time vehicle lane-change prediction and recommendation system on an embedded edge computing platform, specifically NVIDIA Jetson Nano. The system integrates YOLOv5n for vehicle detection and combined with Hough Transform for lane detection using a hybrid parallel-sequential pipeline. The extracted spatial features are seamlessly integrated to construct a Bird’s Eye View environmental representation, which serves as the primary basis for the system to evaluate surrounding traffic conditions and generate preventive maneuver recommendations. To address computational constraints of edge computing platforms without hardware accelerators, an extensive evaluation was conducted across five input resolutions (240p, 360p, 480p, 720p, and 1080p), two power profiles (MAXN and 5W), and an inference optimization strategy. The hardware bottleneck analysis revealed that high resolutions cause severe CPU overloads, leading to GPU starvation and higher latency. Experimental results showed that deploying the system using MAXN power mode combined with Yolo-skip 1 optimization at 360p resolution yields the most optimal configuration. This setup achieves a stable processing speed of 31.58 FPS and an End-to-End latency of 131.24 ms, satisfying the real-time constraints while maintaining reliable functionality for safe driving assistance. Future work will focus on integrating hardware-level accelerators such as TensorRT to further alleviate CPU bottlenecks and fine-tuning the detection model with localized datasets to enhance overall system robustness.