The extraction of target objects of interest from satellite-constellation remote sensing images faces several challenges，such as complex backgrounds，large differences in object scales，and dense target distributions，which significantly limits the practical application of detection algorithms in remote sensing tasks.In recent years，deep learning technology has developed rapidly，and single shot multi-box detector (SSD) algorithms have demonstrated significant advantages in remote sensing object recognition tasks.However，achieving high-precision and low-latency detection on resource-constrained embedded platforms remains a critical issue.In view of this，a remote sensing object detection method based on an improved YOLOv5 framework is proposed in this paper，tailored for embedded deployment scenarios.A lightweight coordinate attention mechanism is introduced to enhance the network’s perception of spatial positional information，effectively suppressing the false positives caused by the background interference.A simplified spatial pyramid pooling module and a bidirectional feature pyramid fusion structure are adopted to improve the multi-scale feature integration.At the post-processing stage，an adaptive non-maximum suppression (Adaptive-NMS) algorithm is designed to reduce false detections by dynamically adjusting the suppression threshold based on the spatial object density.The experimental results show that the proposed method significantly improves the detection accuracy of small objects in remote sensing images while maintaining a lightweight structure.It achieves a mean average precision (mAP) of 85.8% on the remote sensing dataset and reaches a real-time inference speed of 160 frames per second (FPS) on the embedded platform，demonstrating excellent real-time performance and operational efficiency.