CMOS event cameras provide ON/OFF events as intensity changes of a pixel along time. Their main advantage over conventional frame cameras is their low inherent latency and high dynamic range, which opens the way for many new applications. Recent commercial CMOS event cameras by Prophesee-Sony and Samsung show the interest in this line of research. Noise and the lack of a large body of deep learning algorithms and applications similar to their conventional frame cameras counterpart are still open issues for CMOS event cameras. This work aims at these two open questions through the design of low noise CMOS event cameras and deep learning algorithms on events to perform spatio-temporal processing and tackle applications as object detection or tracking.