Research on improvement strategy of DETR real-time object detection algorithm for small devices

With the rapid development of deep learning technology, object detection, as an important task in the field of computer vision, has also been widely applied in the ROS(Robot Operating System) robot field. ROS, as an open-source robot operating platform, provides a favorable operating environment for object detection algorithms. This article studies the improvement strategy of real-time object detection algorithm (RT-DETR) for small devices, which provides a better solution for the field of robotics. Through the application analysis of the RT-DETR algorithm, we found that there are certain issues with detection accuracy and efficiency, as well as the ability to detect small objects, when the performance of small devices is limited. Especially when the camera is not in focus or the image is ghosted, the issues of detection accuracy and efficiency are particularly prominent. For this purpose, we propose a dynamic and irregular deformable convolution kernel strategy to address the performance issues of small edge devices in terms of detection accuracy. In response to efficiency concerns, we propose an enhanced non-linear network structure to achieve greater non-linear capability with fewer parameters applied during the operation process. Finally, we combine the two methods to form a DenNet network (Deformable and Enhanced Nonlinear Convolutional Kernel Networks). Through experimental verification, our improvement strategy can greatly improve the detection accuracy and efficiency of small devices, solve the efficiency problem of insufficient performance of small devices, and has important practical application value.