Money laundering in cryptocurrency transactions is differentiated from traditional financial crimes due to its strong anonymity and decentralization. The existing anti-money laundering techniques cannot be directly applied to cryptocurrency transactions. Considering the traceability, interpretability, and measurability of money laundering crime forensics, this paper designs a four-stage money laundering detection approach: (1) defining a set of transactions of a user in a period as a transaction behavior; (2) constructing a set of features to characterize transaction behaviors; (3) adopting outlier detection and small cluster detection methods to find out loud and subtle anomalous transactions; (4) analyzing the suspicious score distributions of users and calculating a suspected-launderer value for each of them. To evaluate the performance of our proposed method, a real-world money laundering dataset is obtained and experimentally evaluated. The experiment results show that our approach obtains 96.02%, 95.05%, 95.83%, and 95.81% accuracy in terms of abnormal transaction behaviors, suspected money launderers, loud abnormal transactions, and subtle abnormal transactions, respectively, all better than benchmark algorithms. Moreover, the carefully-designed features of transaction behaviors can offer supportive interpretations for the detection results and help exchange security officers to carry on further investigations and crime forensics.