Obstacle Detection for Power Transmission Line Based on Deep Learning

Power line inspection robots often work in the mountains, surrounded by trees and other background disturbances, while the field illumination changes significantly. The distance between the camera and the obstacles has a great influence on the scale changes of the obstacles in the image space. Therefore, accurate obstacle detection is very difficult. The deep learning methods such as Single Shot Multi-Box Detector (SSD) algorithm can be insensitive to illumination and scale changes in complex background with a much high accurate detection. But the amount of parameters of the SSD are so huge that make it very difficult to transplant to embedded systems. Aiming at the problem of stable and accurate detection of obstacles and easy to transplant to embedded systems, this paper proposes a method which significantly reduces the model parameters and improves the detection speed. In this method, the 23 convolution layers of the original SSD are simplified to 7 layers, and the Batch Normalization layer is added after each convolution layer to normalize the convolutional data. The result shows that the algorithm not only ensures the detection accuracy, but also greatly reduces the parameter quantities of the model and improves the detection speed significantly from 4.5 fps of original SSD to 15 fps of simplified SSD on Jetson TX2(an embedded system). Compared with some classical computer vision based detection algorithms, the method is more adaptable to complex environments.