基于机器视觉的雨雾天驾驶辅助系统设计

Design of Driving Assistance System under Rain and Haze Weather Based on Machine Vision

摘要:

随着私家车数量日益增多，雨雾天交通安全问题成为了亟待解决的难题。在嵌入式硬件资源有限的情况下，为驾驶用户设计了基于机器视觉的雨雾天语音辅助驾驶系统。系统结合了湿度传感器、轻量化去雾神经网络AOD-NET和目标检测模型YOLOv5n。在目标检测模型YOLOv5n上，利用K-means++算法重新设计锚框,选取较优的骨干网络并利用模型剪枝进一步压缩模型大小。实验结果表明，改进的模型在Jetson nano上的FPS达到了17.78，最终mAP在人工加雾、分辨率变化的TT100K (Tsinghua-Tencent 100K)数据集到达了65.8%，满足了正常天气与雨雾天气下的驾驶辅助实际应用。

Abstract:

With the increasing number of private cars, traffic safety issues in rainy and foggy days have become an urgent problem to be solved. A voice assisted driving system for driving users in rainy and foggy weather based on machine vision is designed, with limited embedded hardware resources. The system combines humidity sensors, lightweight dehazing neural network AOD-NET, and object detection model YOLOv5n. On the object detection model YOLOv5n, K-means++ algorithm is used to redesign the anchor frame. A better backbone network and model pruning is selected to further compress the model size. The experimental results show that the FPS of the improved model on Jetson nano reached 17.78, and the final mAP reached 65.8% on the artificially fogged and resolution changed TT100K (Tsinghua-Tencent 100K) dataset, meeting the practical application of driving assistance in normal weather and rainy and foggy weather.