Generalized self-cueing real-time attention scheduling with intermittent inspection and image resizing

Abstract

Abstract This paper proposes a generalized self-cueing real-time attention scheduling framework for DNN-based visual machine perception pipelines on resource-limited embedded platforms. Self-cueing means we identify subframe-level regions of interest in a scene internally by exploiting temporal correlations among successive video frames as opposed to externally via a cueing sensor. One limitation of our original self-cueing-and-inspection strategy (Liu et al. in Proceedings of the 28th IEEE real-time and embedded technology and applications symposium (RTAS), 2022b) lies in its lack of computational efficiency under high workloads, like busy traffic scenarios where a large number of objects are identified and separately inspected. We extend the conference publication by integrating image resizing with intermittent inspection and task batching in attention scheduling. The extension enhances the original algorithm by accelerating the processing of large objects by reducing their resolution at the cost of only a negligible degradation in accuracy, thereby achieving a higher overall object inspection throughput. After extracting partial regions around objects of interest, using an optical flow-based tracking algorithm, we allocate computation resources (i.e. DNN inspection) to them in a criticality-aware manner using a generalized batched proportional balancing algorithm (GBPB), to minimize a concept of generalized system uncertainty. It saves computational resources by inspecting low-priority regions intermittently at low frequencies and inspecting large objects at low resolutions. We implement the system on an NVIDIA Jetson Xavier platform and extensively evaluate its performance using a real-world driving dataset from Waymo. The proposed GBPB algorithm consistently outperforms the previous BPB algorithm that only uses intermittent inspection and a set of baselines. The performance gain of GBPB is larger in facing more significant resource constraints (i.e., lower sampling intervals or busy traffic scenarios) because its multi-dimensional scheduling strategy achieves better resource allocation of machine perception.