Research On Road Foreign Object Detection Method Based On Depth Image

Foreign Object Debris(FOD)has long posed a threat to the safety of aircraft operations and the well-being of crew members and passengers.Therefore,developing a system that can detect FOD on airport runways quickly and accurately is of great significance.This thesis proposes two detection models to meet the demand for FOD detection on airport runways.Firstly,a multi-scale detection model is developed that combines road depth images and object detection technology,which is suitable for high-precision requirements and utilizes a fusion and split attention mechanism.Secondly,an improved YOLOv5-based detection model is introduced that features high detection speed.As collecting data on real airport runways is difficult,this thesis conducts FOD detection research on ordinary roads instead.Firstly,in response to the problem of missing values in depth images caused by laser triangulation,this thesis proposes a method for locating and repairing missing values combining adaptivethreshold binarization and an image restoration algorithm based on the fastmarching method.Subsequently,an analysis was conducted on the impact of the collection device vibration.This thesis proposes a convolution method based on a longitudinal kernel to correct the vibration data,which reduces the impact of device vibration on depth imaging.Then,the depth image is processed using a spatial domain piecewise linear enhancement method to enhance the features of road foreign objects in the depth image.Finally,the data set of road foreign object depth images is augmented to 4234 images using rotation and duck-filling methods to enrich the diversity of the data set.Secondly,to improve the accuracy of road foreign object detection,this thesis proposes a multi-scale road foreign object detection model that incorporates a fusion-split attention mechanism.The model is an improved version of the two-stage detection network Cascade RCNN,which adds a Split-Attention mechanism to the backbone network and increases the network depth to enhance the model’s feature extraction capabilities.The original feature pyramid is improved to a PAN+FPN structure,increasing the utilization of different feature layers.The network’s original pooling method is optimized to use ROI Align with bilinear interpolation to avoid losing the features of small foreign object.Experimental results show that the improved network achieves an m AP@0.5 of 95.2% in the road foreign object detection model,which is a 4.6% improvement over the original Cascade R-CNN.Finally,to meet the speed requirements of road foreign object detection vehicles,this thesis proposes a fast road foreign object detection model based on improved YOLOv5.Firstly,the GhostNet network is used to lightweight the YOLOv5 network structure,reducing the parameter and computational complexity by one third.Based on the original feature fusion structure,jump connections are added and redundant nodes are removed to fuse different feature layers by weighting.The original spatial pyramid pooling structure is improved to a serial input structure,which improves the computational efficiency without changing the pooling results.Experimental results show that the improved YOLOv5 can achieve a detection speed of 29.5frames per second,with the highest detection speed reaching 54.4km/h,while maintaining an average precision of 94.4%.In summary,this thesis has completed the acquisition and processing optimization of deep images of road foreign object.Two road foreign object detection models were proposed,one with high accuracy and the other with high detection speed.The research systematically studies the road data acquisition,data processing and foreign body detection,which has application value and reference significance in the field of road foreign object detection.