Research On Key Technologies Of Assisted Driving Detection System Based On Deep Learning

With the increasingly development of social economy,the number of vehicles in family keeps growing.And the problem of driving accident and traffic congestion have become more and more serious.So it is very significant to develop the intelligent assistant driving system in order to reduce the risk of potential traffic dangers.Realtime perception of the road environment is a very critical part of the intelligent assisted driving system.How to recognize,detect,and measure road targets such as vehicles or pedestrians quickly and effectively has always been a research hotspot.Traditional target detection and vehicle ranging methods have some disadvantages such as low detection accuracy,poor environmental adaptability,and high hardware costs.The artificial intelligence technology has been developed rapidly,which injects lots of energy into the target detection and ranging technology based on driving images.However,there are some difficulties on research,such as the complex driving environment,variable light background,and partial occlusion of targets.So the recognition,detection,and ranging algorithms for vehicles and pedestrians need to be improved.Therefore,in order to perceive the enviroment effectively,it is very significant for intelligent assisted driving detection system to study the key technologies of assisted driving detection system based on deep learning.This article introduces the applications and technologies of target detection and ranging in intelligent assisted driving detection system,and analyzes target detection and depth estimation methods based on deep learning.And we studied the key technologies for assisted driving detection from the aspects of target detection and vehicle ranging.Here are the main research content and innovations:1.In response to the topics of target detection and ranging algorithm in the intelligent driving assistance system,we analyze the advantages and disadvantages of traditional methods and the current deep learning methods,propose an assisted driving detection system based on deep learning,and design the system framework and deep learning model.We select a self-supervised learning monocular depth estimation model for ranging application,train it on the KITTI data set,and generate a depth map for the actual driving scene.Combined with the target detection frame,the distance of the preceding vehicle during driving can be estimated easily.2.The deep learning-based one-stage object detection algorithm has a good detection speed and meet the real-time requirements of intelligent driving assistance systems,but there are deficiencies in the detection accuracy.Therefore,we have studied YOLOv3 which was used widely,introduced the self-attention mechanism in Transformer,and embeded the self-attention module in the high layers of YOLOv3 network to capture more global information.At the same time,considering that the local information has more influence on object detection,a Gaussian mask is introduced in the self-attention module to increase the attention score of nearby locations,and the degree of attention score adjustment is learning in train process.The improved model was trained in the MS COCO 2017 data set.Compared with YOLOv3,its mAP@0.5 increased by 2.56%,which effectively improved the detection accuracy.Although there is a little loss in detection speed,detection accuracy is significantly improved when apply to assisted driving system.In this paper,we have focused on the object detection and depth estimation problem,optimized the algorithm performance and verified in practice.And the experiment results have shown that the effectiveness of our deep learning method.In conclusion,our work has theoretical value and practical significance.