遠距離物体検出に適したネットワークアーキテクチャ探索手法
- Author
- 平川翼, 山下隆義, 藤吉弘亘
- Publication
- 自動車技術会春季大会, 2020
Download: PDF (Japanese)
Object detection is an important task in autonomous driving. Especially, detecting not only near but also distant objects in a driving scene is crucial for safety driving and further development of autonomous driving applications such as traffic light recognition and path predictions of pedestrians. While deep learning-based object detection methods now becomes a common approach, the architecture of these methods are manually designed and developed. In this paper, we propose a method to find an optimal network architecture for distant object detections. Our method is based on a neural architecture search (NAS), and the method searches an optimal architecture during training automatically. The experimental results with onboard vehicle camera image dataset for pedestrian detection show that our found optimal network architecture successfully detect distant objects.