Field Robot Environment Sensing Technology Based on TensorRT

The inference speed of complex deep learning networks on embedded platforms of mobile robots is low, and it is difficult to meet actual application requirements, especially in complex environments such as the wild. This experiment out motion blur processing on the data set to improve the robustness, by using NVIDIA inference accelerator TensorRT to optimize the operation, the computational efficiency of the model is improved, and the inference acceleration of the deep learning model on the mobile quadruped robot platform is realized. The experimental results show that, on the test data set, the method achieves 91.67% mAP of 640 x 640 model on the embedded platform Nvidia Jetson Xavier NX. The reasoning speed is about 2.5 times faster than before, reaching 35 FPS, which provides support for the real-time application of mobile robot environment sensing ability in the field.