Obstacle Detection of Pump Truck Based on Knowledge Distillation

Pump truck is an indispensable machine for infrastructure construction. In order to speed up its intelligentization and solve the obstacle avoidance problem of pump truck boom operation, this paper proposes a YOLO-V5 (You Only Look Once-Version 5), Obstacle detection of pump truck based on knowledge distillation, ODPTKD. Firstly, ODPTKD changed the random affine transformation at the input end into new data augmentation, and improved the network structure of YOLO-v5. Then, the temperature coefficient T is used to soften teachers’ output knowledge, and the cross entropy between the students and teachers’ networks is used to get the loss function. Finally, we will use the gradient descent to conduct knowledge distillation and update the parameters of the student model. According to experimental results, MAP of ODPTKD model is 5.1% higher than that of the model without distillation, and the network size is only 34.1% of that of teachers’ network. Compared with the model without data augmentation, the MAP of ODPTKD model is 15.2% higher. What’s more, the ODPTKD model effectively improves the accuracy of target detection, with high real-time frame rate, small memory occupied by the model, favorable for embedded deployment and strong robustness to noise and other interferences.