A object detection method with dynamic real-time constraints
Bo Yu, Xiaoxiao Li, Xiaohui Duan · 2023
Object detection techniques have proven their usability in many conventional domains. However, in driverless vehicles, the task of object detection faces new challenges: Driverless vehicles are highly sensitive to real-time performance, but existing object detection methods do not have the ability to sense and comply with real-time constraints in the real world. In addition, most of the existing models are static and cannot make dynamic trade-offs between efficiency and accuracy based on real-time constraints. Based on the shortcomings of the existing techniques, this paper proposes a dynamic network based on Faster R-CNN, which adds short-circuit blocks and routing units with lower computational costs to the existing models, enabling the network to select the appropriate blocks to execute based on real-time constraints. In addition, the network contains a dynamic RPN module that dynamically adjusts the number of proposals to achieve a dynamic balance between efficiency and accuracy.