Research On Target Detection And Location Based On Deep Learning

Intelligent vehicles consist of environment perception,decision planning,and execution structures.Intelligent vehicles can run normally on the road that the environment perception part plays the most important role.It influences the decisions made at the decision-making layer,and ultimately affects the movement state of the vehicle and the comfort of the passengers.In the environment perception part,the detection and location of each target object in the environment where the vehicle is currently located is very important,and it affects the decision-making of acceleration,braking and overtaking of the vehicle decision-making layer.The traditional object detection and location methods have large false detection rate and missed detection rate,and the location of the object is also greatly biased.Because of the great advantages of deep learning in image detection and location,there is a great theoretical and practical value for deep learning in this area.The common image-based object location and detection system is divided into three steps:image segmentation,object key feature extraction,and object classification.Based on the current requirements for image processing in the intelligent vehicle’s environment perception part,this project conducted detailed investigations on the knowledge involved in each step and the specific performance requirements.Deep learning has a wide range of applications,so it also corresponds to different deep learning models.In the area of ? ? image processing explored in this project,many improvements and optimization models are based on the most basic model of Convolutional Neural Networks(CNN).The convolutional neural network can build a deep network structure,can make the network reach deep,can abstract higher-level abstract features of the image,and make more efficient use of image information.After analyzing the detection and location of the object in the image by the VGG network,this paper builds an improved VGG model based on the VGG network model,effectively integrates the three steps of conventional image-based object detection and location,and improves the model.Call it IVGG.After studying the implementation principle of the region proposal network(RPN)for image segmentation,an improved strategy was proposed to improve the performance of the RPN network.The improved RPN network is applied to the IVGG model built on this topic.In the IVGG model constructed this time,space pyramid pooling technology(SPP)is introduced to improve the detection accuracy and recognition probability of the target object.Using the current open source Caffe deep learning development tool to build the proposed model,and using the current KITTI dataset commonly used in the field of intelligent vehicles to train and test the model built,according to the corresponding test results to assess the accuracy of the model built.At the end of the article,the difficulties encountered in the learning process of this topic and relevant solutions are summarized.At the same time,relevant future prospects for intelligent vehicle detection and recognition of image processing are discussed.