Autonomous driving network under multi-scale perception and global planning

Abstract

The invention discloses an autonomous driving network under multi-scale perception and global planning, mainly focuses on deeper exploration of perception capability and effective improvement of obstacle avoidance planning, network interpretation and other capabilities, and provides an effective solution for the three weak core problems. Firstly, the network is structured and layered to improve the interpretation capability of each part of the network; secondly, the multi-scale perception capability of the network is improved by pre-training an FPN network [15], an obstacle avoidance module isadded to optimize the vehicle planning capability, and finally, the flow direction is controlled by perceiving obstacle avoidance features and planning a navigation instruction, so that effective feedback of the vehicle to global intention and local features is realized.

Description

technical field [0001] The invention relates to the technical field of fault analysis and safety assurance in the navigation and positioning process of an unmanned autonomous driving system, and in particular provides an autonomous driving network under multi-scale perception and global planning. Background technique [0002] Since Barret Electrnies of the United States developed the first self-driving car in 1950, self-driving technology has become the hottest research direction in the automotive field. After Google joined the self-driving camp in 2009, self-driving technology quickly became a powerful bargaining chip for major high-tech companies to compete in technology, and successively launched Waymo, Cruise, Apollo, Zoox, Nuro and other outstanding self-driving cars [1] . These autonomous driving systems are uniformly designed according to the perception, planning and control modules, and are tuned through a large number of cloud simulations. Although these leading a...

AI Technical Summary

Problems solved by technology

[0004] (1) Lack of overall planning ability. When driving to the intersection, the driving direction is not determined, and the corresponding behavior cannot be completed according to the driving intention inside the system. It will only reproduce the driving action of the trained driver after driving to the current intersection as much as possible. make corresponding choices;

[0005] (2) Perception, planning, and control layers are strongly coupled together, and the design is too black-box, so that it is impossible to return to the source of the error based on the error sample, and it is impossible to optimize the layered algorithm by modifying the effects of each layer;

[0006] (3) The perception ability is too weak. When the network model is deconvoluted, it can be seen that the learned feature areas are concentrated on the boundaries of the road and the surrounding larger objects, but ignore the structured target information such as vehicles and pedestrians that should be paid more attention to. [7] ;

[0007] (4) Lack of obstacle avoidance ability, it is difficult for the network to learn the braking or steering behavior of obstacle avoidance

Although the realization of this network largely solves the problems of learning singleness and lack of global planning of self-driving vehicles, etc. [4] , but there is still no effective solution to the problem (2-4). The autonomous driving network still lacks interpretability, lacks learning ability for obstacle avoidance problems, lacks multi-scale structured perception ability, and still focuses on the road for learning features. Boundary and lack of learning methods for important structured objects (cars, pedestrians, etc.)

Images