Active millimeter wave imaging (AMWI) is an efficient way to detect dangerous objects concealed under clothes. However, because the images acquired by AMWI are often obscure and some of concealed objects are small in size, the automatic detection and localization of the objects remain as a challenging problem. Yao[1] first employed convolutional neural networks (CNNs) and used a dense sliding window method to detect concealed objects. In this paper, we make two improvements on Yao"s work: (1) Using contextual information to suppress interference and improve detection rate; (2) Using a two-step search method instead of exhaustive search to reduce computing complexity. We firstly use one CNN in vertical direction to filter the interference and get the vertical position of the concealed object, then use another CNN to determine the horizontal position. To make use of big window containing contextual information, we use IoG (intersection-over-ground-truth) instead of IoU (Intersection-over-Union) to define positive and negative samples in training and testing process. Experimental results show that our proposed method reduce the computing time to about 30% while achieving better detection performance.