Research On Lane And Traffic Object Detection Algorithm Based On Convolutional Neural Network

With the acceleration of traffic intelligence,vehicle intelligence has become the focus of attention.Accurate lane detection and object detection are the basis for vehicles to realize functions such as lane departure warning and automated emergency braking.Convolutional neural networks can automatically obtain corresponding feature extraction capabilities by learning a large amount of data.Compared with traditional feature extraction methods,convolutional neural networks have better generalization capabilities in the face of complex scenes.Therefore,it is of great significance to develop accurate and robust lane detection and object detection algorithms based on convolutional neural networks to promote the development of autonomous driving technology.Based on the convolutional neural network,a two-stage lane detection algorithm is proposed.The first stage uses a semantic segmentation network to extract the lane pixels in the image,and the second stage performs instance segmentation on the extracted lane pixels.In the first stage of algorithm design,a feature association block is proposed for the slenderness of the lanes,which enhances the flow of feature information in the network and improves the network’s reasoning ability for lanes.The dual channel upsampling block increases the gradient propagation path and improves the convergence speed of the network.In the second-stage algorithm design,the DBSCAN algorithm is improved to dynamically adjust the neighborhood,which can effectively improve the instance segmentation effect of lane post-processing algorithm.Finally,through the experimental test,it is proved that the proposed lane detection algorithm has a good accuracy.Combining the research results of lane detection with yolov5 object detection network,a multi-task network for traffic scene is proposed.The network improves the utilization of network parameters by matching single encoder and multi decoder,and can detect lane and traffic object at the same time.Compared with the detection method of two single-task networks in series,the multi-task network improves the detection speed by 22.84% and has better real-time performance.In addition,a multi-task datasets was made for campus traffic scene.The dataset contains 5000 images,with a total of40494 traffic objects and 14350 lanes marked.In order to test the performance of the multi-task network in practical application scenarios,a vehicle collision avoidance system is designed by combining multi-task networks with a variety of monocular vision algorithms.The proposed collision avoidance system is tested on the on-line controlled vehicle platform,which proves the effectiveness of the system and the practical value of the multi-task network.