Condar: Context-Aware Distributed Dynamic Object Detection on Radar Data

Deep neural networks are an established class of algorithms widely used for real-time data analysis in computer vision-based mobile applications. However, on one hand the constrained resources - computing power, energy reservoir and memory - available to mobile devices clash with the complexity of high-performance neural models. On the other hand, offloading the execution of the computing task to a wirelessly connected edge/cloud server requires the transfer of information rich signals over volatile and capacity constrained channels. To address this issue, the Split Computing (SC) paradigm has been proposed. In SC frameworks, a baseline DNN model is used to create two models, namely head and tail models, that are executed at a mobile device and edge/cloud server, respectively. In this work, a SC paradigm - Condar - is presented that introduces several core design and conceptual innovations: (i) we target object detection on radar data rather than 2D RGB images; (ii) we design encoders/decoder structures that are specialized to the specific operating environment; (iii) we develop a multi-branched model that can dynamically select the best performing encoder while executing the main model. Our design is based on the CenterNet model with a split ResNet50 backbone trained on the RADIATE dataset’s radar images of resolution 1152x1152. Our results show that Condar has performance analogous to that of a Generalized CenterNet Model with a difference of 4.41%, 1.72%, and 2.18% on mAP@50, mAP@75, and mAP@COCO respectively, while simultaneously performing compression on the features of the data to 15.5KB. This datasize is smaller than lowest possible compression data size of JPEG compression at 21KB, which results in a considerable performance degradation.