Research On Freeway Spill Detection Based On Improved Neural Network

With the development of the scale of the freeway network,the number of vehicles driving into the freeway is also increasing,coupled with the characteristics of the freeway limit speed faster compared to ordinary highway,freeway accidents also occur frequently.Among all kinds of major accidents,freeway spills are also an important cause of freeway accidents.Whether it is a tiny spill or a cargo scattered by freight vehicles,it always endangers the safety of vehicles on the freeway,so the current real-time accurate detection of freeway spills has become a key problem to be solved urgently.At present,a lot of research on neural network is mainly in pedestrian detection,face detection,text detection,traffic signal and remote sensing target detection,etc.,while the research on traffic signal detection is usually conducted in urban environment,the research on traffic detection in freeway environment is relatively less,and the research on freeway spill detection is even much more.Therefore,it is important to study the freeway spill detection algorithm.In this paper,based on the existing target detection research,we propose a CA-YOLOv5-based detection model to directly improve the detection accuracy of the model;from the perspective of feature fusion,we propose a BIFPN-YOLOv5-based detection model to improve efficiency of the model by improving the efficiency of feature fusion,and then improve detection accuracy of the model.For the problem of misdetection of spilled objects,combined with the relationship between spilled objects and vehicles,this paper proposes a re-detection mechanism of spilled objects based on IOU comparison to reduce the chance of misdetection of vehicles carrying goods,bumpers and other items as spilled objects.In summary,the main work of this paper is summarized in the following five aspects:(1)To address the problem that the spill dataset is difficult to obtain,this paper collects and labels the spill dataset by means of network collection.To increase the data volume as well as to ensure the data sample diversity,this paper adopts the restricted contrast adaptive histogram equalization,Laplace operator,logarithmic transform,Retinex SSR,Retinex MSR,Retinex MSRCR image enhancement algorithms,and rotation and cropping image geometric transformation methods to enhance and expand the dataset.(2)To address the problem of difficult detection of throwing objects,this paper proposes a detection model based on attention mechanism from the perspective of detection mode,and structurally proposes two structures with attention mechanism placed after SPPF layer and attention mechanism combined with CSP as CSPA,and combines the two structures with SE,CA,and CBAM attention,respectively,and by using the COCO128 dataset for experiments yields the the optimal model is the detection model based on CA-YOLOv5,which improves the detection accuracy of the model.(3)For the problem of difficult detection of throwing objects,this paper also proposes a detection model based on BIFPN-YOLOv5 from the perspective of feature fusion.In terms of structure,the BIFPN network with stronger feature fusion capability is used to replace the original PAN network,which further enhances the feature fusion capability of the model and thus improves the detection accuracy of the model.Finally,the effectiveness of the improvement is verified by training experiments using the COCO128 dataset.(4)To address the problem of easy misdetection of spill,a spill re-detection mechanism based on IOU comparison is proposed in combination with the relationship between spill and vehicles,and a spill detection process based on IOU comparison is proposed according to the target detection process,and it is proved in the subsequent spill detection experiments that the mechanism can effectively reduce the chance of the detection model misdetecting the vehicle carrying goods,bumpers and other items as spill.(5)Above all,the CA-YOLOv5-based detection model is combined with the BIFPNYOLOv5-based detection model,and the CABIFPN-YOLOv5-based freeway spill detection model is proposed.In the freeway spill detection experiments,the PASCAL VOC2012 dataset is used to train the model considering the data volume of the homemade spill dataset,and the training weights are used in the spill detection experiments by the migration learning method.The effectiveness of the CABIFPN-YOLOv5-based spill target detection model is verified by comparative analysis.Based on the above work,CABIFPN-YOLOv5-based detection model’s mean average precision m AP@0.5 reaches 85.1 and m AP@0.5:0.95 reaches 73.2 under spill dataset training condition,with FPS of 79 frames,can meet the demand of real-time detection,while combining with the IOU comparison-based spill inspection process effectively solves the problem of spill being difficult to detect and the problem that vehicle cargo is easily misdetected as spill.