In this work, we utilize the framework of compressed sensing (CS) for distributed device detection and resource allocation in large-scale machine-to-machine (M2M) communication networks. The devices deployed in the network are partitioned into clusters according to some pre-defined criteria. Moreover, the devices in each cluster are assigned a unique signature of a particular design that can be used to indicate their active status to the network. The proposed scheme in this work mainly consists of two essential steps: (i) The base station (BS) detects the active clusters and the number of active devices in each cluster using a novel block sketching algorithm, and then assigns a certain amount of resources accordingly. (ii) Each active device detects its ranking among all the active devices in its cluster using an enhanced greedy algorithm and accesses the corresponding resource for transmission based on the ranking. By exploiting the correlation in the device behaviors and the sparsity in the activation pattern of the M2M devices, the device detection problem is thus tackled as a CS support recovery procedure for a particular binary block-sparse signal $x\in\mathbb{B}^N$ -- with block sparsity $K_B$ and in-block sparsity $K_I$ over block size $d$. Theoretical analysis shows that the activation pattern of the M2M devices can be reliably reconstructed within an acquisition time of $\mathcal{O}(\max\{K_B\log N, K_BK_I\log d\})$, which achieves a better scaling and less computational complexity of $\mathcal{O}(N(K_I^2+\log N))$ compared with standard CS algorithms. Moreover, extensive simulations confirm the robustness of the proposed scheme in the detection process, especially in terms of higher detection probability and reduced access delay when compared with conventional schemes like LTE random access (RA) procedure and classic cluster-based access approaches.