Research And Embedded Implementation Of Infrared Small Target Detection Algorithm For UAV

The rapid development of infrared imaging technology makes it occupy an irreplaceable position in many civil and military fields,such as field search and rescue,urban fire control,ecological protection,early warning guidance and so on.Therefore,infrared small target detection has attracted extensive attention in the field of infrared target detection.The infrared small target detection method based on traditional image algorithm uses low-level feature information for analysis and detection,resulting in a high false alarm rate in infrared small target detection under complex background.Deep learning technology can obtain high-level semantic information,which is more efficient than traditional image algorithms.Therefore,infrared small target detection based on deep learning method is a research hotspot and has great practical significance.The low signal-to-noise ratio of the infrared image and the incomplete target structure result in the inconspicuous contour and low contrast of the infrared target.At the same time,due to the small target occupying less pixels,the deep convolution neural network is easy to cause the loss of feature information,resulting in poor boundary box regression effect and a large number of false positive prediction results.To address the above problems,this paper focuses on the requirements of UAV airborne computer manifold 2G scene,and takes the convolutional neural network target detector as the starting point to carry out the following research work:Aiming at the problem of insufficient infrared small target datasets,a new dataset named Infra Tiny is constructed by using M100 UAV and Zenmuse XT infrared camera.The dataset contains 3218 images and a total of 20893 bounding boxes.There are 17896 bounding boxes with 32×32 pixels,accounting for about 0.856 of the total.Target categories include person and car.Aiming at the problem of inability to obtain multi-scale key information and loss of feature information in infrared small target detection,an infrared small target detection model named Infra-yolo is designed.The model integrates the proposed multi-scale attention mechanism module(MSAM)and feature fusion augmentation pyramid module(FFAFPM)based on yolov3.The MSAM enables the network to obtain scale perception information by acquiring different receptive fields,while the background noise information is suppressed to enhance feature extraction ability.The FFAFPM aims to enrich semantic information,enhance the fusion of shallow feature and deep feature,and optimize the regression task of the network.Finally,to enable Infra-yolo run on the UAV embedded devices,the channel pruning method is adopted to compress the Infra-yolo model,and the knowledge distillation is used to improve the performance of the pruned model.In actual deployment,Tensor RT quantization technology is used to further accelerate the inference of the pruned model.In this paper,a large number of experiments are carried out on Infra Tiny dataset.The experimental results show that Infra-yolo can significantly improve the detection performance of infrared small targets.Infra-yolo is 2.7 percentage points higher than yolov3 and 2.5 percentage points higher than yolov4 on m AP.When the learning parameters of Infra-yolo are reduced by 88%,the pruned model is 0.5 percentage points higher than yolov4 on m AP.After quantifying with FP16,the inference time of pruned model on Manifold2 G is reduced from 171 ms to 68 ms,which greatly improves the inference speed of the model on embedded devices.