Detection of the objects using various models and technologies is performed within the domain of computer vision. This has an application in the medical industry, automation, and self-driving car industries, among many others. It is predicted that by the year 2025 there will be over 8 million self-driving cars running on roads. According to studies, by 2050 almost all cars and public transportation will be automated. Object detection is an essential and one of the most important tasks for self-driving cars. It allows the vehicle to sense and navigate their environment in a safe and efficient manner. This involves identifying and locating a variety of things in the car's field of vision including traffic lights, road signals, pedestrians and other cars. The goal of this research is to investigate and evaluate the issues that arise when autonomous automobiles use their visual subsystem to identify things. This paper proposes a unique custom Detection Transformer model for visual object detection in autonomous vehicles, specifically addressing the limitations of pre-trained models such as reduced accuracy for small objects and slower processing times. The study incorporates a self-attention mechanism that occurs after the DETR backbone has extracted features. Using an attention mechanism in the model will enhance the DETR's performance on the custom dataset. The proposed model achieves a 20% improvement in accuracy for detecting small objects compared to pre-trained models and can detect up to 100 objects per second with over 90% accuracy. These findings suggest that the proposed model can significantly enhance the performance of object detection applications in real-world scenarios.
TransVOD: Transformer-Based Visual Object Detection for Self-Driving Cars
08.05.2024
1136761 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Self-Driving Cars Perception Pedestrian Detection
Springer Verlag | 2024
|Springer Verlag | 2023
|