Benchmarking of Embedded Object Detection in Optical and RADAR Scenes

A portable, real-time vital sign estimation protoype is developed using neural network- based localization, multi-object tracking, and embedded processing optimizations. The system estimates heart and respiration rates of multiple subjects using directional of arrival techniques on RADAR data. This system is useful in many civilian and military applications including search and rescue.

The primary contribution from this work is the implementation and benchmarking of neural networks for real time detection and localization on various systems including the testing of eight neural networks on a discrete GPU and Jetson Xavier devices. Mean average precision (mAP) and inference speed benchmarks were performed. We have shown fast and accurate detection and tracking using synthetic and real RADAR data.

Another major contribution is the quantification of the relationship between neural network mAP performance and data augmentations. As an example, we focused on image and video compression methods, such as JPEG, WebP, H264, and H265. The results show WebP at a quantization level of 50 and H265 at a constant rate factor of 30 provide the best balance between compression and acceptable mAP.

Other minor contributions are achieved in enhancing the functionality of the real-time prototype system. This includes the implementation and benchmarking of neural network op- timizations, such as quantization and pruning. Furthermore, an appearance-based synthetic RADAR and real RADAR datasets are developed. The latter contains simultaneous optical and RADAR data capture and cross-modal labels. Finally, multi-object tracking methods are benchmarked and a support vector machine is utilized for cross-modal association.

In summary, the implementation, benchmarking, and optimization of methods for detection and tracking helped create a real-time vital sign system on a low-profile embedded device. Additionally, this work established a relationship between compression methods and different neural networks for optimal file compression and network performance. Finally, methods for RADAR and optical data collection and cross-modal association are implemented.

final draft