Research On Multi-object Detection And Segmentation Of Traffic Scene Via Deep Learning

Unmanned driving is an important technology to ensure traffic safety and improve comfortable driving.Multi-object detection and segmentation technology of traffic scenes plays a key role in understanding the unmanned driving scenes.In traffic scenes,pedestrians,vehicles,traffic signs and traffic lights that appear any time will be more difficult to detect and segment due to the factors such as occlusion,uneven illumination and bad weather.In order to ensure the accuracy,robustness and real-time requirements of multi-object detection and segmentation,this paper uses computer vision and deep learning to focus on the improvement of multi-object detection and image segmentation algorithms in traffic scenes.On this basis,we completes a detection and segmentation intelligent system that can be equipped with a mobile terminal.The main research contents are as follows:Summarize the categories involved in multi-object in traffic scenes,and introduce the classic datasets used to train the improved network.Discuss and analyze common classic methods and deep learning-based methods of object detection and image segmentation,so that we can extract preliminary plans for how to carry out multi-object detection and segmentation.Analyzes the deep learning framework and operating mechanism of Caffe and Tensor Flow for algorithm deployment.To solve the problems of poor effect and slow detection speed for multi-object detection methods in traffic scenes,an improved Faster R-CNN algorithm is designed.By improving the Ro I network layer structure,the characteristic information of small objects in traffic scenes can be completely transmitted between the high and low layers in the network to solve the problem of low robustness for small object detection.Replacing the two quantization algorithms by the bilinear interpolation algorithm makes the internal feature aggregation of the network become a continuous process,which effectively improves the algorithm computing power and reduces the time loss of the algorithm.The domestic and foreign open source datasets and self-made datasets are used as the overall data to ensure that the improved algorithm model can effectively learn the object characteristics.The improved multi-object detection model obtained by adjusting the hyperparameters shows good rapidity and robustness in the subsequent open source dataset test experiments.Aiming at the problems of low accuracy and poor robustness for multi-object image segmentation,an improved Mask R-CNN multi-object rapid segmentation method in traffic scenes is designed.First,the lightweight Mobile Net is used as the backbone network to effectively reduce the internal parameters and the model volume,which improves the transplantation ability of subsequent mobile terminal algorithm.Second,by optimizing the FPN network and the backbone network convolution calculation method,the object feature information can be effectively used in the pyramid to ensure the accurate segmentation of small objects in the image.Replace the normalization function and adjust the hyperparameters to obtain a multi-object segmentation model of traffic scenes.In the experimental verification,the improved model can achieve good results on internationally recognized datasets such as Berkeley BDD-100 K,Baidu Apollo Scape and Autonomous driving Nu Scenes datasets.The average detection accuracy can reach 88.2% and the detection speed of a single image can reach 0.402 seconds,which can adapt to a variety of complex traffic scenes to complete rapid segmentation of multiple objectives.Based on the actual requirements of mobile multi-object detection and image segmentation in traffic scenes,an intelligent system for multi-object detection and segmentation is designed.The hardware mobile platform and software working modules of the system are designed,meanwhile,the working process and different operating modes of the system are introduced.In several intelligent system evaluation experiments,the system completes multi-object detection and segmentation with high accuracy under the premise of ensuring real-time performance,which lays a foundation of theoretical algorithm and practical application for multi-object detection and segmentation of traffic scenes in intelligent driving and unmanned driving.