With　the　development　of　smart　cities,　visible　light　communication　(VLC)　with　its　unique　advantages　is　increasingly　regarded　as　a　viable　complement　to　traditional　radio　frequency-based　wireless　communications.　In　practical　applications,　line-of-sight　VLC　is　susceptible　to　blocking/shadowing,　resulting　in　communication　interruptions.　Even　though　non-line-of-sight　(NLOS)　transmission　can　effectively　address　this　issue,　propagating　signals　are　often　subject　to　significant　attenuation　and　multipath　effects,　which　can　degrade　the　quality　of　communications.　In　this　paper,　we　propose　a　NLOS　VLC　system　with　chirp　spread　spectrum　modulation,　which　leverages　reflected　light　to　overcome　blocking.　Additionally,　a　spatial　shift　convolutional　neural　networks　(S2-CNN)　demodulator　is　used　to　mitigate　the　signal　linear　and　nonlinear　transmission　impairments　introduced　in　NLOS　propagation,　thus　achieving　effective　joint　signal　compensation　and　recovery.　Experimental　results　demonstrate　that,　S2-CNN-based　demodulator　can　effectively　compensate　for　linear　and　nonlinear　distortions,　achieving　a　transmission　rate　of　more　than　10　Mbps　over　a　2.7-m　NLOS　link,　demonstrating　higher　reliability　and　robustness.　©　1983-2012　IEEE.