Pedestrian Detection And Distance Estimation Of Underground Roadway Based On Deep Learning

In an unmanned system,computer vision is often used to complete road detection and object detection on the road.The road detection includes the lane and the detection of the travelable area;the detection of road signs on the road includes the detection and classification of all traffic participants such as detection of other vehicles,pedestrian detection,traffic signs and signal detection.At the same time,the target is detected for ranging.However,in the underground mining and transportation process,the front rail and the target on the track are also required to detect the underground unmanned system.The research in this paper is for the detection and ranging of pedestrians in the underground roadway.The main contents are as follows:(1)A Dense-YOLO pedestrian detection algorithm for underground roadway is proposed.First,the k-means clustering strategy is used to calculate the size of the anchor box.Secondly,in order to make better use of the upper layer or the first few layers,it is proposed to extract network features by using dense connection blocks and reuse "collective knowledge" to enhance feature learning propagation.Then,the non-maximum suppression algorithm is used to suppress the non-maximum value of the multiple candidate frames generated at the position of the same target in the detection process,and the candidate frame with the highest score is selected.Finally,network model training,verification and test experiments were performed on the data collected underground.During the test,it was found that for pedestrians with small distances in the image,accurate detection could not be achieved.To this end,a network structure combining multi-level feature maps was proposed,which can be applied to shallow feature maps,deep and shallow.The layer features are combined to precisely locate small targets.The experimental results show that the proposed algorithm has a detection accuracy of 93% and a speed of 25 frames per second,realizing the detection of real-time and efficient roadway pedestrians.(3)Aiming at the problem of distance estimation of pedestrians in front of underground locomotives,a method of monocular vision ranging is proposed.First,the camera’s interior/exterior parameters are obtained by camera calibration,and distortion correction is performed at the same time.Then,the monocular ranging model is built,and the relationship between the human and vehicle distance is calculated by the coordinate system conversion.Finally,the distance calculation is performed by combining the coordinate parameters obtained by the pedestrian detection algorithm.The experimental results show that the calculation error of the monocular ranging method is less than 5%,which proves the applicability of the ranging method.Finally,the paper summarizes the overall content of the paper and looks forward to the next step of the research.