Optimizing One-Stage Object Detection for URLLC: A Two-Stage Pipeline with Oculi SPU

Pedestrian-related accidents account for an estimated 20-25% of the approximately 1.19 million road fatalities occurring annually, highlighting the urgent need for enhanced detection systems in surveillance and smart transportation to mitigate such incidents. The escalating demand for Ultra-Reliable Low-Latency Communications (URLLC) in modern surveillance and transportation systems has become a solution to that problem but highlighted significant challenges in reducing inference times of machine learning models, particularly on constrained hardware. The maximum permissible latency for most vehicular applications ranges from 6 to 20 milliseconds based on the application. When tested on an 8GB RAM setup using the NTU, LLVIP, and VIRAT datasets, YOLOv8n achieved a minimum inference time of approximately 200 milliseconds per frame, making it unsuitable for URLLC applications despite 8GB RAM being an adequate hardware setup for edge devices. This research introduces a new two-stage pipeline that integrates Oculi's Sensing and Processing Unit (SPU) simulation with the YOLOv8n one-stage object detection model to enhance pedestrian detection, thereby addressing these challenges. By leveraging Regions of Interest (ROIs) generated by Oculi's SPU, our approach optimizes the input data for the YOLO model, substantially reducing computational overhead and thus making it suitable for URLLC applications. This integration can lead to a reduction in inference time by up to 93% in the tested frames when being compared to YOLOv8n pre-trained model.