To address error accumulation and limited adaptability to complex road conditions in traditional modular intelligent-driving architectures, this study proposes an end-to-end decision-making algorithm based on a convolutional LSTM network and a self-attention memory module. Incorporating human driving logic, the basic model integrates spatiotemporal feature extraction, historical state modeling, and longitudinal and lateral control. Improved residual units and depth wise separable convolution enhance feature representation efficiency, while future-sequence prediction supports task optimization. A self-attention memory module is further introduced to strengthen long-term temporal and spatial modeling, forming an optimized end-to-end decision model. Simulation results show a 100% success rate and 0% collision rate in simple scenarios, and a 94% success rate with a 4% collision rate in complex scenarios. Prediction accuracy reaches 88.7% for steering and 89.1% for speed. Driving trajectory, heading angle, and steering wheel angle closely match expert driving behavior. The proposed method also demonstrates good ride smoothness and real-time decision-making performance, confirming its effectiveness in improving the accuracy, robustness, and safety of autonomous driving and providing useful technical support for advanced end-to-end autonomous driving systems. The above experimental results are only based on the CARLA simulation platform Town4 scenario, which is constrained by the simulation scenario. The generalization ability of the model in more complex urban scenarios needs further verification.
