Research On Vehicle Detection Algorithm In Real Traffic Scene Based On Deep Learning

With the rapid economic development,China’s vehicle ownership data has increased exponentially,and the continuous increase in vehicle ownership has also brought a series of traffic safety problems,which seriously affects people’s quality of life.This makes real-time vehicle detection a way to predict traffic in advance.Utility tool for security issues.And with the rapid development of deep learning technology,it continues to change the field of vehicle detection.This paper applies the deep learning model to real-time vehicle detection in surveillance video scenarios.The main research contents of this paper are as follows:(1)Vehicle detection based on feature fusion SSD.As an earlier single-stage detection algorithm with excellent results,the first content of this article is the SSD-based vehicle detection algorithm.In view of the problem of the original SSD model’s inability to detect small targets,the SSD network model is analyzed.Targeted improvements include learning from the residual module to deepen the neural network,avoiding over-fitting and disappearing network gradients,adding fast fusion to the basic structure of the SSD model,and making full use of the detailed information of the shallow features and the semantic information of the deep features.Comparing the original algorithm with the improved algorithm,it can be concluded that the improved SSD algorithm has better detection performance than the original SSD algorithm.At the same time,the algorithm has excellent robustness and environmental adaptability in a complex traffic environment.It has anti-interference ability under bad weather conditions.(2)Improved YOLOv3 vehicle detection with embedded the hole convolution module.Compared with the SSD algorithm,YOLOv3 has made many improvements.First,the SSD algorithm classifier is Softmax,which can only output one category,while YOLOv3 can output multiple categories at the same time,and the detection effect is improved compared with SSD,so the second The algorithm uses the YOLOv3 algorithm.In view of YOLOv3’s weak ability to detect small target vehicles and occluded vehicles,this paper proposes an improved YOLOv3 vehicle detection algorithm.Firstly,a cavity convolution module is embedded in the original and feature extraction network Darknet-53 to reduce the loss of target information The receptive field is enhanced,and the traditional NMS algorithm is improved in this paper.If the Io U of the prediction frame is greater than the set threshold,it will be attenuated in a certain way.The algorithm can obtain accuracy and real-time performance at the same time,which is conducive to the realization of fast and accurate automatic vehicle detection.(3)Vehicle detection based on Corner Net.Both SSD and YOLOv3 are essentially based on the anchor box method to classify and regress vehicle targets.However,there are two problems with vehicle detection based on anchor frames: firstly,there are too many anchor frames,which easily leads to extreme imbalance of positive and negative samples,and secondly,the anchor frame mechanism will bring many parameter processing problems,which is time-consuming to process.It is longer,so this paper adopts the key point-based vehicle detection method Corner Net,which detects the vehicle target bounding box as a pair of key points,namely a pair of upper left corner and lower right corner,so there is no need to design an anchor frame,thereby solving the anchor frame The two problems brought about by the mechanism play a more important role in the field of target detection..