Real-Time Embedded Object Detection System Design Based on Deep Neural Networks

With the rapid development of the Internet of Things (IoT) and edge computing technologies, achieving high-accuracy real-time object detection on resource-constrained embedded devices has become a significant challenge in the field of computer vision. This paper proposes a real-time embedded object detection system design based on deep neural networks. The system is built upon a lightweight YOLO network architecture and integrates depthwise separable convolution, attention mechanisms, and multi-scale feature fusion techniques to construct an efficient detection model suitable for embedded platforms. Through model compression techniques such as 8-bit integer quantization and structured pruning, the storage requirements and computational complexity are significantly reduced. Additionally, a hardware acceleration strategy based on the ARM NEON instruction set is designed to fully exploit the parallel computing capabilities of multi-core processors. Experimental results on the COCO2017 dataset demonstrate that the proposed system achieves a real-time detection speed of 25.4 fps on the ARM Cortex-A78 processor, with an mAP@0.5 of 89.3%. Compared to the baseline method, inference speed is improved by 206%, model size is reduced by 79.6%, and power consumption is only 2.8W. This system provides an efficient and practical object detection solution for applications such as intelligent surveillance, autonomous driving, and robotic navigation, and holds significant application value on mobile and edge computing devices.