Dynamic Multi-ROI Parallel Inference Architecture for Online Video

Computer vision technology is crucial for environmental perception in autonomous driving, but online computer vision tasks based on object detection often need to perform object detection task first and then downstream tasks, which consumes a lot of time. This paper proposes a dynamic multi-ROI parallel inference architecture for online video analysis, which uses the correlation between video frames to parallelize object detection and downstream tasks, which greatly improves the execution efficiency of the algorithm. Based on this architecture, a two-step object detection algorithm based on parallel inference architecture is further evolved through model sharing, which effectively improves the accuracy of small object detection in high-definition video. The method proposed in this paper is not only suitable for autonomous driving tasks, but can also be extended to more online video analysis scenarios. A large number of experimental data prove that the parallel inference architecture has a significant efficiency improvement effect for online video analysis based on object detection, and can effectively improve the accuracy of small object detection in specific scenes of high-definition video.