Multi-task Deep Learning Network In Autopilot Scene

In recent years,machine learning technology has made continuous breakthroughs,especially deep learning technology has often made more than expected progress in different fields.As an important field of deep learning application,autopilot technology has attracted much attention.Deep learning technology is generally used in object detection,semantic segmentation,object tracking,depth estimation and other functions.The neural network which takes into account two or more functions at the same time is called multi-task neural network.In practice,it is found that the biggest contradiction between the limited computing force of the vehicle chip and the huge computing consumption of the deep model is to block the landing of the application.A single-function deep neural network is difficult to run on a vehicle chip in real time,not to mention the need to run different functional neural networks at the same time in order to provide secure autopilot.Therefore,with the deepening of autopilot technology research and development,multi-task deep learning neural network has become one of the important solutions of autopilot technology landing.Based on the analysis of objects in autopilot scene,a new multi-task deep learning neural network model is proposed in this paper.Based on the similarity between the underlying dimensional boundary of the entity target and the segmentation boundary of different targets in the dense convex target graph,the model uses the object detection method to deal with the segmentation problem by disassembling the drivable areas into road objects.The new model can not only deal with the tasks of object detection and region segmentation,but also cancel the coding and decoding structure of the previous multitasking network,reduce the model structure and operation flow,and improve the real-time performance of the multitasking network.In addition,the multi-objective frame fusion method avoids the possible misjudgment and interference noise in the key areas of decision-making,and improves the robustness of the model.Based on the analysis of the problem of object depth estimation,a new object depth estimation algorithm is proposed in this paper.The algorithm obtains the structured understanding of the vehicle target through the clustering analysis of the rear view appearance of the vehicle.On this basis,the traditional image method is improved so that it can be applied to multi-task neural networks.In this paper,the effectiveness and stability of the algorithm are verified in theoretical calculation and practical site test.The comparison and analysis of the traditional image and depth estimation segmentation network methods show that the algorithm is more accurate.In this paper,the algorithm is integrated into the multi-task deep learning network,so that the model can deal with a variety of tasks at the same time,such as object classification,object detection,drivable areas segmentation and depth estimation.