Direct time-of-flight (D-ToF) image sensors based on Time-Correlated Single Photon Counting (TCSPC) systems face two main challenges, i.e., the limitation of spatial resolution by the amount of data to store and the accuracy of the reconstruction under high background noise. The key contribution of this paper to overcome these issues consists in the introduction of a custom pixel-wise Compressive Sensing (CS) hardware implementation in combination with a deep learning reconstruction algorithm. Besides the reduction of the pixel pitch, the CS approach limits the probability of counters overflows, enabling larger photon counts operating mode, easing the depth/intensity reconstruction process in practice. Compared to prior work on deep learning models for TCSPC data, our proposed approach achieves a similar depth-intensity reconstruction accuracy in the typical lowphoton flux mode of operation. However, when combined with the proposed CS hardware implementation compatible with high photon counting, our solution outperforms the most advanced SPAD sensing strategies as well as the best-in-class remote processing, both in terms of intensity-depth reconstruction performance and pixel pitch reduction.