Cognitive Map Lane Layer Generation System For Urban Intelligent Vehicle

With the deepening of autonomous driving research,the industry generally believes that the realization of autonomous driving above L4 level requires the support of dedicated maps.High-precision maps are a viable solution,but their high cost has led to slow progress.In response to this problem,the team proposed a lightweight,low-cost map for autonomous driving system-cognitive map.The cognitive map is represented by a layer structure,where the lane layer is the basis of the cognitive map.The research work of this paper is centered on the cognitive map lane layer generation system for urban intelligent vehicle,and mainly completes the following three tasks:(1)Expression and generation of the lane layer of the cognitive map.i)In view of the high cost of high-precision maps,this paper designs the expression of the lane level of the cognitive map according to the cognitive process of human driving vehicles to achieve the same function at a lower cost.ii)In view of the problem of low efficiency of manual mapping,this paper designs a method for generating the intersection part of the cognitive map lane layer and a method for generating the non-intersection part of the cognitive map lane layer,which improves the mapping efficiency by 70%.(2)Intersection recognition method combining semantic segmentation and shape classification.The intersection feature is an important basis for generating the intersection part of the lane layer of the cognitive map.In view of the problem of lack of intersection recognition methods,this paper proposes an intersection recognition network that combines semantic segmentation and shape classification.The network uses a splittransform-merge strategy to achieve a feature expression effect close to a complex network with a more lightweight structure,and introduces The multi-level layer jump structure makes the network focus on the global features to improve the recognition rate.(3)Based on Lanenet improved road shoulder line detection method.Road shoulder or lane line features are an important basis for generating non-intersection parts of the lane layer of the cognitive map.In view of the lack of efficient and accurate road shoulder and lane line detection methods,this paper proposes an improved detection method.This method splits the end-to-end deep learning network into multiple small modules,and connects each module with traditional machine learning methods.The accuracy and speed of the method in the campus dataset are increased by 5.4% and 17%,respectively,and the problem of zebra crossing misdetection and the difficulty of migration training convergence are solved.This paper designs the expression and generation method of the lane layer of the cognitive map,and proposes the intersection recognition method and the improved road shoulder and lane line detection method,and finally successfully builds a cognitive map lane layer generation system for urban intelligent vehicle,which improves Cartographic efficiency.Provides a technical foundation of intelligent vehicle for applications in city scenarios.