This paper proposes a vehicle state classification model utilizing drones for achieving highly reliable autonomous driving vehicles. The model offers a wide field of view, providing improved data for the autonomous driving assistance system by considering the relationships with existing vehicles in the video. The object detection model, based on Convolutional Neural Networks, is trained using image data collected by drones. The YOLOv5 model is employed for detecting vehicles on the road. Detected vehicles are cropped into five candidate regions based on proximity, which serve as input data for the vehicle state classification model. The proposed model introduces a depthwise CSP block to effectively learn high-density layers and facilitate efficient weight learning. Additionally, a method that leverages deformable residual block to set receptive fields relying on features extracted from the fixed receptive field of the conventional convolutional kernel is proposed, enabling faster identification of critical features. Experiments are conducted using directly collected drone data to evaluate the performance of the trained model and compare it with a model based on dense layers. The approach aims to provide a more reliable and efficient vehicle state model for a robust autonomous driving system.
DE-VSNet: Dense Efficient Vehicle State Classification Using Drone Traffic Dataset
2023-08-09
2324506 byte
Conference paper
Electronic Resource
English
City-ATM – Safe Drone Operations in Dense Traffic
IEEE | 2021
|CQSkyEyeX: A Drone Dataset of Vehicle Trajectory on Chinese Expressways
Springer Verlag | 2024
|