A pedestrian target detection method for underground coal mine based on image fusion and improved CornerNet-Squeeze

In unmanned driving and security monitoring in the coal mine, detecting pedestrian targets is very important. But under the influence of special working conditions such as dim light, uneven illumination, complex background, and small and dense pedestrian targets, the pedestrian targets in the image have some problems such as few edge details, low signal-to-noise ratio and high similarity with the background. It is difficult to effectively identify the pedestrian targets under multi-scale occlusion. In order to solve the above problems, a pedestrian detection method for underground coal mine based on image fusion and improved CornerNet-Squeeze is proposed. The image collected by the infrared camera and depth camera is fused at the pixel level using the two-scale image fusion (TIF) algorithm. The morphological processing is carried out for the fused imoge to reduce background interference. Based on the CornerNet-Squeeze network, octave convolution (OctConv) is introduced into the hourglass type backbone network to process the high and low frequency information of image features, so as to enhance the image edge features and improve the detection capability of multi-scale pedestrians. The experimental results show the following points. ① The improved CornerNet-Squeeze model can effectively improve the detection precision of underground pedestrian while maintaining the real-time performance of the original algorithm on the data sets of range image, infrared image and fusion image. ② The detection precision of the model trained by the fusion image dataset is higher than that of the models trained by the infrared image dataset or the depth image dataset. The result shows that the fusion image can give full play to the advantages of the depth image and the infrared image, and is helpful to improve the detection precision of the model. ③ In the six scenes of different degrees of occlusion and multi-scale pedestrian target, the model trained by the improved CornerNet-Squeeze has the lowest pedestrian misdetection rate. ④ Compared with YOLOv 4, the average accuracy of the improved CornerNet-Squeeze algorithm on the COCO2014 pedestrian dataset is improved by 1.1%, and the detection speed is improved by 6.7%. ⑤ The improved CornerNet-Squeeze can effectively detect the small target in the image. The detection capability of the small target is obviously improved.