Research on safety evaluation and decision algorithm of automatic driving based on artificial intelligence

This paper constructs a multi-level and multi-dimensional safety evaluation index system for self-driving cars, covering five aspects: environmental awareness, decision-making and planning ability, vehicle control ability, system reliability and compliance with laws and regulations. Based on AI technology, it designs a safety evaluation model with five levels: data collection, processing, feature extraction, safety evaluation and feedback optimization. Secondly, aiming at the shortcomings of the existing decision-making algorithms, an automatic driving decision-making algorithm based on deep reinforcement learning (DRL) is proposed, which adopts an improved deep Q-network (DQN) framework and combines dynamic reward mechanism and state attention mechanism to improve the decision-making ability in complex traffic scenes. The experiment uses CARLA simulation platform, designs a variety of complex scenes, and verifies the effectiveness of the proposed algorithm. The results show that DRL-Attention algorithm is significantly superior to traditional DQN, PPO and rule-based algorithms in key indicators such as collision rate, average vehicle speed and times of sudden acceleration and deceleration. At the same time, the key roles of attention mechanism and dynamic reward mechanism in improving safety and perception ability are verified by ablation experiments.