Low-altitude flyer detection method based on compressed SSD

Abstract

The invention provides a low-altitude flyer detection method based on a compressed SSD, and relates to the technical field of target detection. The channel pruning algorithm is adopted to compress the model, and the detection speed of the model is increased. The method comprises the following steps: firstly, adding a batch normalization layer in a basic model, and introducing a scale factor for each channel; then, a scale factor is jointly trained, so that the numerical value of the scale factor is associated with the importance of the channel. And finally, judging the importance of the channels by measuring the size of the scale factor, and cutting off the channels with low importance. The performance of each aspect of the compressed model is superior to that of the original model, the network structure of the model is more refined, the size is greatly reduced, the operation speed is higher, and a higher detection speed can be obtained. According to the channel pruning method adopted by the invention, the importance of each channel is measured, so that the influence of the pruned channels on the precision of the model is small, and therefore, the precision of the optimized model is close to that of the original model.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to a low-altitude flying object detection method based on compressed SSD. Background technique [0002] Low-altitude airspace refers to the flight area with an altitude of less than 1000m, which has important application value for agricultural production, medical assistance and transportation industries. In recent years, in order to support the domestic aviation industry, my country has gradually opened up low-altitude airspace, and its management problems have also followed. At present, the common low-altitude flying objects are mainly birds and drones. As a high-tech product, drones are widely used in various fields, such as agricultural production, travel aerial photography and competitive competitions. It brings convenience and happiness to people, but also has potential safety hazards. Some privately owned drones will pose a safety threat to passenger aircraft flying...

AI Technical Summary

Problems solved by technology

Among the existing target detection models, the SSD (Single ShotMultiBox Detector) model performs well in terms of detection speed and accuracy, but the SSD model is complex in structure and bulky, and its detection speed is relatively slow for low-altitude flying objects

Images