Hybrid AI Models for Real-Time Object Detection in Low-Bandwidth Environments

Abstract

The demand for real-time object detection in constrained environments such as remote surveillance, autonomous navigation, and low-power edge devices has surged significantly. However, achieving high accuracy and responsiveness in low-bandwidth settings remains a substantial challenge due to the high computational cost of deep learning models and the inability to transmit high-resolution data. This paper explores a hybrid AI framework that integrates lightweight Convolutional Neural Networks (CNNs), rule-based filters, and adaptive compression techniques to achieve optimal object detection performance. The proposed architecture performs feature extraction at the edge using a compressed model while leveraging cloud-based heavy models selectively through a hybrid decision layer. This dual-layer strategy minimizes data transmission overhead, balances latency and accuracy, and enables effective inference under connectivity constraints.

Extensive simulation research is conducted using standard datasets (e.g., PASCAL VOC, COCO Lite) under emulated bandwidth limitations (e.g., 128 kbps to 512 kbps). The hybrid model demonstrates up to a 43% improvement in detection accuracy compared to standalone lightweight models and reduces inference latency by 28% relative to full-cloud inference. Statistical analysis confirms the significance of these results. This manuscript contributes a viable solution for deploying intelligent object detection systems in bandwidth-scarce applications, bridging the gap between edge and cloud AI.