An improved scheme based on feature fusion for the lightweight object detection algorithm deployed in embedded devices is proposed in this paper. The algorithm is based on the MobileNetV2-SSD lightweight object detection algorithm with the idea of FPN feature fusion. It combines the three feature layers of MobileNetV2-SSD that contain more semantic information, and is regenerated based on the fused feature layer Feature pyramid for object detection. The improved model and the original model are compared and tested on the PASCAL VOC data set. Compared with the original model, the mAP is increased by 3.6%, and reaches 76.5%. At the same time, the detection effect of small targets is significantly improved. Finally, it is tested in the embedded device Jetson AGX Xavier. The improved model has a detection speed close to the original model when the network structure is more complex, reaching the detection speed of 22FPS, which can realize real-time object detection on embedded devices.