Location-based services (LBS) are witnessing a rise in popularity owing to their key features of delivering powerful and personalized digital experiences. The recent developments in wireless sensing techniques make the realization of device-free localization (DFL) feasible in wireless sensor networks. The DFL is an emerging technology that utilizes radio signal information for detecting and positioning a passive target while the target is not equipped with a wireless device. However, determining the characteristics of the massive raw signals and extracting meaningful discriminative features relevant to the localization are highly intricate tasks. Thus, deep learning (DL) techniques can be utilized to address the DFL problem due to their unprecedented performance gains in many practical problems. In this direction, we propose a DFL framework consists of multiple convolutional neural network (CNN) layers along with autoencoders based on the restricted Boltzmann machines (RBM) to construct a convolutional deep belief network (CDBN) for features recognition and extracting. Each layer has stochastic pooling to sample down the feature map and reduced the dimensions of the required data for precise localization. The proposed framework is validated using real experimental dataset. The results show that our algorithm can achieve a high accuracy of 98% with reduced data dimensions and low signal-to-noise ratios (SNRs).