Method, system, computer equipment and storage medium for unmanned platform smoke and dust perception

Abstract

The present application relates to a method, system, computer equipment and storage medium for sensing smoke and dust on an unmanned platform. The method comprises: constructing a training data set by sparsely labeling the smoke and dust image, training the smoke and dust recognition network, generating a feature activation image according to the recognition result, calculating the activation distribution of neurons for the smoke and dust and background features, and analyzing the training data according to the activation distribution. set for supplementary labeling. Define the loss function including the smoke and dust classification loss item and the feature activation loss item, iteratively update the training data set for the neural network, and the unmanned platform performs smoke and dust perception based on the trained smoke and fog perception model. This application is based on the classification network architecture, and according to the response relationship between the neural network and the background or smoke features in the smoke image, the sample is marked with pixels, and the training process of the neural network is supervised, which reduces the complexity of the model and the cost of labeling, and enhances the The ability to rapidly deploy models.

Description

technical field [0001] The present application relates to the technical field of unmanned platform smoke, dust and fog perception, in particular to a method, system, computer equipment and storage medium for unmanned platform smoke, dust and fog perception. Background technique [0002] The sensors configured on unmanned platforms mainly include visible light cameras, millimeter-wave radars, and lidars. Lidars have become the main sensors of most current unmanned platform perception systems due to their more accurate measurement capabilities and stronger night-time adaptability. However, due to the short wavelength of laser light emitted by lidar, it cannot penetrate fine particles in the environment, so it is extremely susceptible to interference from smoke and dust, causing the environmental perception system to fail, and ultimately affecting the normal operation of ground unmanned platforms under smoke and fog conditions. Therefore, to ensure the stable operation of the u...

AI Technical Summary

Problems solved by technology

The smoke and dust target detection technology is mainly based on the improvement of the target detection network, such as the YOLO series and the SSD series, whose structure is more complex than the recognition network, and the amount of calculation and the labeling burden of training samples are relatively large

The above-mentioned solutions are all overly dependent on the quality of feature design, and their ability to deal with complex environments is poor, and the designed features often fail because they cannot take into account all situations

Images