Research On Classification And Detection Of Objects On The Road Based On Deep Learning

With the rapid economic development,the number of cars has increased sharply,and the road infrastructure is not yet perfect,which has caused frequent accidents.In order to improve this situation and prevent traffic accidents,the intelligent assistant driving system and intelligent braking system developed by the researchers have produced good effects in some case.The classification and detection of target objects are the key to these intelligent systems.The traditional image classification and target detection technology takes a long time to detect objects and the accuracy is not high enough,which affects the development of smart cars.Based on the advantages of deep learning technology in image classification and object detection,the paper has made a deeper research on the classification and detection of road objects.According to the basic concept of Convolution Neural Network(CNN),the paper divides the basic structure of the model into five parts:input layer,convolutional layer,pooling layer,fully connected layer,and output layer,and reconstructs CNN.The model conducts identificationand classification research on 11 types of vehicles such as ambulances,buses,and taxis on the road,and at the same time,the identification model is dynamically enhanced and expanded.Through the optimization of the convolution layer through the Inception Module,the optimization and solution of the model found that the network depth of the model was shallower and the convergence speed became faster.At last,the accuracy rate of 90.12% was obtained on the data set,which is correct compared to the transfer learning.Increased by nearly 4.21%.For the widely used target detection frameworks of Faster RCNN,SSD and Yolov3,the paper uses the KITTI dataset to train these three object detection frameworks,and compares their mean average precision(mAP)and transmission frame rate per second(Frames Per Second,FPS),the paper selects Yolov3 as the object detection reference frame of the paper.To optimized for the framework,this paper took three measures.1)To balance the categories by increasing the number of objects of fewer categories.2)For the disadvantage of inaccurate target positioning,this paper readjusted the value of anchor in different data sets.3)For the difficult detection and missed detection of small targets,this paper added a scale fusion.At last,he mAP value of the optimized Yolov3 object detection framework reaches 0.7912,which is higher than the original Yolov3 mAP.It has increased by nearly 5.31%,and the detection speed has also reached the real-time detection requirements.It has highpractical value and provides useful experience for research and detection of other objects.