The burgeoning field of Artificial Intelligence for the Internet of Things (AIoT) is rapidly transforming smart transportation, posing the formidable challenge of efficiently processing voluminous traffic data. Addressing this, our study presents a novel three-tiered device-edge-cloud video stream scheduling framework, DECS-DRL, underpinned by deep reinforcement learning. The DECS-DRL framework enables efficient video stream processing through multi-layer collaboration: key frame extraction, using lightweight models for fast task processing, and using more complex models for high-precision task processing. Central to navigating the intricate balance between detection accuracy and latency minimization in task scheduling, we introduce an innovative deep reinforcement learning algorithm enhanced by a hybrid state encoder. Our empirical evaluations underscore the DECS-DRL framework’s efficacy, showcasing significant advancements in video streaming task scheduling and optimizing deep reinforcement learning model training within intelligent transportation systems.