Object Detection In Traffic Scene Based On Deep Learning

In recent years,the rapid development of computer technology and artificial intelligence has promoted the research of intelligent traffic.Intelligent transportation system which depends on intelligence and information is also an important research direction of intelligent transportation in the future.Among them,traffic signal light detection technology based on deep learning is one of the key technologies,which has important research significance and application value.YOLOv3 is a deep learning framework commonly used in target detection in recent years.In this paper,by improving the network structure of YOLOv3,it is further studied to be used in the detection scene of small target traffic signal lights,and a target detection model incorporating image restoration preprocessing module is designed to improve the detection accuracy of small target traffic signal lights in rainy days.The main work completed in this paper is as follows:Firstly,a dense-yolov3 target detection model with Dense connection is proposed.Based on the analysis of the working principle and network architecture of DenseNet,a dense-yolov3 target detection model is proposed,which integrates DenseNet and YOLOv3.In this model,the idea of dense connection is introduced to enhance the feature reuse in the propagation process of the network and realize the feature fusion between different layers,so as to simplify the feature extraction network and improve the detection effect of small targets.On this basis,K-means clustering algorithm is used to calculate the anchor frame scale suitable for small target traffic signal lights,and GIoU is introduced to replace IoU in YOLOv3 model to optimize network performance.The experimental verification of this model on Bosch data set and Lara data set shows that the dense-Yolov3 target detection model with Dense connection is more accurate than the original Yolov3 model in the target detection accuracy of traffic signal lights.Then,a traffic signal detection model based on image restoration preprocessing under rainy conditions is proposed.On the basis of analyzing the working principle of the attention mechanism,an attention mechanism module is added to focus the attention on the target to be detected to increase the accuracy of target detection.A CBAM-Dense-YOLOv3 target detection model fused with attention mechanism is proposed.On this basis,the image restoration preprocessing operation is introduced,the heavy rain restoration model is used as the pre-algorithm of the traffic signal target detection network to construct a neural network model,and a traffic signal detection model based on image restoration preprocessing under rainy conditions is proposed.Then,the traffic signal detection model based on image restoration preprocessing is used to carry out experimental verification on the synthetic rainwater effect data set.The results show that the proposed model has certain advantages in the traffic signal detection under rainy conditions.Finally,a traffic signal detection system based on deep learning is designed and implemented.Based on the PyQt5 framework,a traffic signal detection system is implemented,which can perform traffic signal detection on uploaded pictures or video files,and synchronously output the detection results in the window.