Behavior decision-making is a core technology for vehicle intelligence. Deep reinforcement learning (DRL), with its environment-interactive capability and end-to-end decision-making advantages, has shown great potential in this field. This paper conducts a multidimensional analysis and systematically reviews the core content and development trends of DRL-based autonomous driving behavior decision-making research. First, the development of behavioral decision-making is reviewed, and the application trends of DRL in autonomous driving is analyzed. Second, a five-dimensional framework “state-action-reward-policy-evaluation” is proposed to analyze the mapping between algorithmic components and driving tasks such as car-following and lane-changing. Third, application schemes of DRL in uncertain environments are examined through typical traffic scenarios including ramp merging, intersections, and construction zones. Finally, we identify key challenges such as multi-vehicle coordination, long-tail event handling, and algorithm interpretability, and suggest future research directions. The study shows that, technically, DRL algorithm selection and optimization are becoming more diverse, with models evolving toward multi-modal and lightweight designs. In terms of application paradigms, behavior decision-making is transitioning from single-vehicle intelligence to vehicle-road-cloud collaboration, and from function-driven implementation to human-centric interaction. Overcoming current technical bottlenecks requires a co-evolution path of algorithm innovation, hardware acceleration, and regulatory adaptation.