Robust Adaptive Modular Perception System (RAMPS): A Node-Based Processing Framework for Training and Testing Vehicle Perception Systems

Neural network-based algorithms have advanced the state-of-the-art in many perception-related tasks such as detection, tracking, and prediction, which are critically important for autonomous vehicles, ADAS, and other related technologies. However, integrating disparate neural network-based algorithms together can be challenging due to varying system-level requirements, non-standard network input and output formats, and custom network features. In this paper, we introduce RAMPS, a framework for developing a node-based system for integrating datasets, perception algorithms, and other processing modules in a single cohesive graph-based system. This system represents a novel framework architecture for constructing and evaluating perception pipelines. We first outline the design philosophy of this system, then we explain the design decisions we made in developing a functioning prototype, then outline algorithms and functions that we have already integrated into RAMPS, and lastly demonstrate some practical examples and benchmarks that show how our proposed framework operates and benefits current and future state-of-the-art perception systems. We intend to release our framework as open source software in the near future for the benefit of the autonomous vehicle, ADAS, and overall computer vision communities.