Research On Pedestrian Detection Technology Of Multi-Information Fusion For Mining Vehicles

At present,mining vehicles are mainly driven by workers.Due to the intricate environment and poor light condition in the underground roadways,transportation accidents,such as vehicles hit people,happen occasionally,causing enormous loss to coal mining enterprises.Therefore,it is necessary for mining vehicles to identify pedestrians at the direction of moving,which is to prevent the accidents happening.For the above problems,a pedestrian detection method is designed for the special environment of underground coal mine.The main contents are as follows:(1)For the special underground environment,image and label enhancement need to be operated on the image data collected and marked in order to expand the training dataset.The method of Clahe histogram equalization is adopted to expand the brightness distribution of the image.The noise in the image is dealt with bilateral filtering,to retain the image feature information and remove noise as much as possible so an underground pedestrian detection dataset can be constructed.(2)For the problem that the traditional YOLOv3 algorithm is difficult to adapt to the special underground environment,a Caps-YOLO algorithm,that is for underground mining vehicle to detect pedestrian,is proposed.Dense connection method is adopted to replace the Residual connection of the original network to construct a Dense Block component,improving the utilization of feature maps.In the classification layer,the Capsule structure combined with the dynamic routing mechanism is adopted to realize the classification function of the model.And Caps-YOLO learns the attribute information of the object from multiple angles,which is to captures various states of the object and effectively improves detection precision for underground pedestrian.Compared with the original YOLOv3 model,the average precision of Caps-YOLO is improved by 3.25%,3.78%,2.69%and 4.94%in multi-object small object,red light and occlusion scenes,respectively.(3)A detection method that combines the advantage information of visible image,infrared image and depth image is adopted.The images of visible and infrared are input to the Caps-YOLO model for training using a joint feature extraction method.Then calculating the depth density value of each pixel in the depth image to classify its pixel into different types.and the probability that each pixel belongs to a different type of object is obtained,which can be used to optimize the Caps-YOLO detection result and improve adaptability of the model to the special underground environment.Compared with the single visible model,the average precision of the fusion multi-information model is improved by 6.90%,and the maximum detection time is 120.41ms,which meets the requirements of precision and real-time of pedestrian detection for mining vehicle.Therefore,the Caps-YOLO algorithm proposed is of great significance to ensure the safety production of coal mine enterprises.