Vehicle detection by satellite videos is important for urban intelligent traffic construction. A satellite videos vehicle detection method is proposed to solve the problems of the surrounding background and the vehicles having a similar color, and the low number of pixels occupied by vehicles. First, we design a weighted normalization sparse autoencoder network, highlighting the effective features and strengthening the expression of the vehicle features by introducing weighted normalization sparse autoencoders into the backbone. Second, we design a method for the early fusion of spatial information transmission, realizing the enhancement of spatial information in the deep feature map through atrous convolution. Finally, we propose K-means++ based on adjusted cosine similarity, stretching the anchor box size by introducing the adjusted cosine similarity calculation. The experimental results show that the average accuracy of this method in satellite videos vehicle detection reached 75.7%, and the recall rate reached 89.6%. This method can effectively detect vehicles in satellite videos, and the effect is significantly improved compared with other methods.