Recently, unmanned aerial vehicles are widely used in surveillance, aerial photography, power grid line inspections and other places. In order to deploy the YOLOv3 [1] algorithm on drones, it is necessary to adopt the YOLOv3 algorithm with fewer parameters and a simpler structure. This paper implements the model compression of YOLOv3 based on methods such as sparseness, pruning, and knowledge distillation. This paper implements the sparseness of the network by adding L1 regular expressions on the convolutional layer. After that, redundant channels and layers are removed through channel pruning and layer pruning. After sparse and pruning, the mAP lost a lot. By using knowledge distillation after pruning, it attempts to recover mAP lost in sparseness and pruning. By this method, the YOLOv3 algorithm can be deployed on embedded platforms such as RK3399pro. We evaluate the model on the visdrone2019 dataset. The experimental results show that after model compression, YOLOv3 is more suitable for deployment on embedded platforms.