Positioning method and system for autonomous driving through long short-term memory (LSTM)-based deep reinforcement learning (DRL)

A positioning method for autonomous driving through long short-term memory (LSTM)-based deep reinforcement learning (DRL), comprising:

• performing normalization preprocessing on a complex environment of autonomous driving based on a Partially Observable Markov Decision Process (POMDP), to acquire a real-time kinematic (RTK) positioning result;
• constructing an LSTM-based DRL model, wherein the LSTM-based DRL model comprises a policy network and a value network, and the policy network comprises an input layer, an LSTM network layer, a hidden layer, and a output layer;
• training the LSTM-based DRL model in the complex environment of a autonomous driving, comprising:
• modeling the complex environment of autonomous driving based on the POMDP to acquire a POMDP environment;
• interacting the LSTM-based DRL model with the POMDP environment to acquire an observation buffer of a sliding window, and
• iteratively updating parameters of the policy network and the value network by the observation buffer of e sliding window;
• inputting the RTK positioning result into the LSTM-based DRL model for correction to acquire an evaluated value of a position correction action;
• and
• sequentially outputting, by the LSTM-based DRL model, corrected position information to a real road in the complex environment of autonomous driving based on the evaluated value of the position correction action.