Advanced driver assistance systems (ADAS) are growing in quantity and complexity, requiring even more sensor effectiveness and computational resources. Even though cameras and LiDAR sensors can provide a steady solution in ideal conditions, they are unreliable for adverse weather conditions, such as rain and dense fog. Therefore, radar technology is used for those circumstances is an option to be used for those circumstances. However, its classification accuracy must be enhanced to compete with the other sensor types. This work presents a solution to enhance radar classification quality, adapting a Convolutional Neural Network (CNN) structure to the input of micro-Doppler images of vulnerable road users (VRUs), such as pedestrians, cyclists, and e-scooter drivers. Moreover, two CNNs were trained, each for a particular region of interest (ROIs), one in the near field of the radar until 15m, and the second one up to 25m. The proposed solution presents a promising overall accuracy for its first trial. Nevertheless, it must be further extended to more complex scenes, including more participants, obstacles, and weather conditions.
Convolutional Neural Network Classification of Vulnerable Road Users based on Micro-Doppler Signatures using an Automotive Radar
2022-10-08
1698232 byte
Conference paper
Electronic Resource
English
British Library Conference Proceedings | 2021
|