| Wires and cables in China are densely distributed,low in height,and are difficult to quickly recognized in the air.With the increasing use of helicopters and unmanned aerial vehicles(UAV),it is easy for such aircraft to collide with wires and cables at a low altitude,resulting in flying accidents.The radar-based method is affected by the resolution of the equipment,so the detection distance is relatively close.The traditional image processing methods need a lot of manual parameters and are not robust enough.In view of such problems,this paper mainly applies a CNN-based multi-source wire detection method applicable to visible and infrared images.And the method is trained and tested in a specific scene.The main work of this paper is as follows:1)Studied the extraction method of local features of wire image.In order to train the wire feature extraction ability of the feedforward convolutional neural network and the full convolutional network,the image slices of small batch and the image slices generated by simulation are used.In the test,the accuracy rate of the feedforward neural network in visible image detection was 86.9%,while the accuracy rate of the method based on u-net in infrared and visible image detection is higher than 90%.2)Studied the method of detecting the possible areas of wires in large-scale images.In this paper,a multi-scale electric wire presence region detection network is proposed.In the test of existing image data slice,the classification accuracy is 99%.Combined with the feature extraction network,the whole wire detection process is realized in two steps.3)Studied the fusion method of visible light and infrared results without retraining the network.When visible light and infrared collecting devices are relatively fixed,Harris corner detection is used to extract feature points,and then matching points are selected manually for matching.Then the local features of the wire and the wire area were fused,and the effect was improved.The detection rate of local features increased from 98.9% to 99.3%,while the accuracy rate of regional detection increased from 99.3% to 99.6%.4)Realized the demonstration system of desktop and embedded platform based on Python voice.The TFLite and Tensor RT libraries were used to optimize the model to meet the harsh operating environment of the embedded platform.Finally,the test results on the embedded platform show that the detection can basically meet the real-time requirements,and the model capability didn’t decrease much. |
