Optimization of Visual AI applications for next-generation networked and distributed edge scenarios is an important and challenging problem area given the computation, power and bandwidth resource constraints of client devices and edge servers. Dynamically adapting to variations in system resource availability and optimizing the trade-offs in accuracy vs. compression rate and computational complexity is important for system efficiency. An emerging paradigm for the deployment of complex Deep Neural Network (DNN) models for video analytics in these edge computing scenarios is split-DNN computing, where the DNN model is partitioned with one part executed on a client device and the other part on an edge server. Earlier work has largely addressed split-DNN computing in the context of image analytics. However, the application to video sequences presents significant challenges of computational complexity. In this paper, we propose a flexible and low-complexity approach to address these specific challenges for distributed DNN-based video analytics and semantics-preserving learned compression. We combine lightweight bottleneck encoder-decoder neural networks for compressing deep feature representations along with optical flow-based warping of these deep features. We demonstrate significant compression gains measured with a BD-Rate of \(-82.68\%\) for object detection and \(-59.31\%\) for segmentation when compared with the earlier image-based analytics and compression approaches, and even larger gains over conventional video compression. In addition, we enable dynamic optimization of split-DNN video analytics at the edge by providing lightweight training and inference approaches with simple solutions for fine-grained adaptation in the complexity-rate-accuracy space.