The integration of artificial intelligence (AI) into portable and embedded systems has led to a paradigm shift in how deep learning models are developed and deployed. In particular, the increasing demand for real-time computer vision tasks on resource-limited hardware has emphasized the need for computationally efficient architectures. Object detection, a cornerstone of computer vision, typically involves high computational complexity, substantial memory requirements, and elevated power consumption—rendering traditional convolutional neural networks (CNNs) impractical for use in low-power environments such as smartphones, Raspberry Pi, and edge IoT platforms. This paper presents an in-depth experimental analysis of lightweight CNN architectures tailored for real-time object detection on constrained devices. We critically examine state-of-the-art models including MobileNet, YOLOv4-Tiny, SqueezeNet, and EfficientDet-Lite, assessing them across diverse performance parameters. Key metrics such as model footprint, mean average precision (mAP), inference latency, frames per second (FPS), and power consumption are used to benchmark these models on representative edge hardware. In addition, we explore the efficacy of optimization techniques such as quantization, pruning, and neural architecture search (NAS) in enhancing model efficiency without significantly compromising detection accuracy. Through rigorous evaluation and comparative analysis, the study highlights the trade-offs between accuracy and computational efficiency, providing practical guidance for model selection in real-world scenarios. Case studies and empirical results offer insight into performance bottlenecks and optimization potential, while a forward-looking discussion addresses ongoing challenges and future research directions. This work aims to bridge the gap between high-performance object detection and resource-aware deployment, paving the way for scalable, energy-efficient AI applications on the edge.

Keywords: Lightweight CNNs, Edge AI, Object Detection, Low-Power Devices, Real-Time Inference, Model Compression