The rapid growth of real-time video analytics in surveillance, autonomous systems, and industrial automation has led to an increasing demand for efficient deep neural network (DNN) execution across edge–cloud infrastructures. Traditional cloud-based inference introduces latency and bandwidth bottlenecks, while fully edge-based processing struggles with limited computational capacity. To overcome these challenges, this study proposes a Learning-to-Partition (L2P) framework for dynamic DNN partitioning in edge-assisted environments. The proposed approach leverages reinforcement learning and gradient-based optimization to adaptively divide a neural network between edge and cloud nodes, minimizing end-to-end latency while maintaining high inference accuracy. Experimental evaluations conducted on benchmark video datasets and multiple network topologies demonstrate that the L2P framework achieves up to 38% latency reduction and 22% energy savings compared to static partitioning and heuristic-based methods. Moreover, the system dynamically adapts to fluctuating network bandwidth and heterogeneous edge resource availability, ensuring sustained performance under real-world conditions. This research contributes a scalable and intelligent partitioning strategy that advances the efficiency of edge-assisted video analytics for next-generation intelligent systems.