As one of the most reliable and discriminative biometrics, fingerprint has been widely applied in various applications such as personal electronic products, secure payments, forensics and security. By comparison with contact-based ones, contactless fingerprint recognition systems can avoid many risks, such as fingerprint contamination, uncontrolled non-linear distortion and hygienic concerns. However, this technology also presents new challenges for contactless automatic fingerprint recognition systems (AFRSs), such as low-contrast ridge-valley, perspective distortion, real-time 3D fingerprint reconstruction and effective 3D fingerprint feature. For example, traditional transformation based methods tend to suffer from inaccurate alignment, due to perspective distortion. Also, start-of-the-art methods cannot achieve real-time 3D fingerprint reconstruction. This dissertation focuses on contactless AFRSs, including 2D contactless fingerprint pre-processing, 2D contactless fingerprint recognition, 3D fingerprint reconstruction, and 3D feature extraction and recognition. The main contributions of this thesis are summarized as follows. 1. Propose an effective framework to pre-process 2D contactless fingerprint images, which aims to extract fingertip regions and strengthen the contrast between ridges and valleys. In the framework, a method based on convolutional neural network is first developed to detect fingertip regions, and then a method based on intrinsic image decomposition and guided image filtering is proposed to enhance the contrast between ridges and valleys in the fingertip region. 2. Propose a robust contactless fingerprint recognition method based on a global minutia topology and a loose genetic algorithm. Compared to existing methods that rely mainly on recovering the transformation between two fingerprints to establish the maximum minutiae correspondence and achieve a comparison score, we formulate fingerprint verification as a combinatorial optimization problem. By optimizing the elaborately defined energy function to directly determine the minutiae correspondence, the proposed method can effectively avoid the distortion and deformation problem inherent in contactless fingerprints. Further, a minutia-pair expanding algorithm is proposed to complement the optimal minutiae correspondence to establish the final correspondence through which the recognition accuracy can be improved. 3. Propose a real-time, low-cost and accurate 3D fingerprint reconstruction method. In this method, a novel ridge-valley-guided approach is proposed to accurately establish the ridge-ridge correspondence between two fingerprints in real time. First, fingerprint images of different views are pre-processed to generate a served ridge-valley map (SRVM). Then, a ridge-valley correspondence matrix (RVCM) is defined to measure the ridge-valley (RV) correspondence degree and calculated by our proposed dynamic algorithm in one-dimensional space. And then, the ridge-valley correspondence between the two SRVMs is established according to the RVCM by our proposed greedy algorithm. Finally, a disparity map is refined by our curve-based smooth operation and is used to reconstruct the 3D depth information. Comprehensive experimental results demonstrate that our method outperforms the state-of-the-art methods in terms of reconstruction time and accuracy. 4. Propose a novel and effective three-dimensional topology polymer (TTP) feature to achieve accurate 3D fingerprint recognition. The TTP feature codes 3D minutiae topology by projecting the 3D minutiae onto multiple planes and extracting their corresponding 2D topologies, which has been proven to be effective and efficient. First, 3D minutiae are extracted and represented in multiple new 3D spaces. Then, the TTP is extracted by projecting the 3D minutiae onto multiple planes and coding their corresponding 2D typologies. Besides, the TTP feature is computational efficiency and memory efficiency, and therefore facilitates the 3D fingerprint recognition. Comprehensive experimental results demonstrate that the proposed method outperforms the state-of-the-art methods in terms of both recognition accuracy and recognition time.