Camera-based　parking　occupancy　detection　driven　by　deep　learning　algorithms　is　a　promising　technique　for　building　the　parking　guidance　and　information　system.　However,　when　the　available　camera　is　looking　at　a　long　parking　lot　with　a　relatively　low　angle,　the　deep　learning　method　will　fail　to　detect　vehicles　accurately　as　vehicles　closer　to　the　camera　will　block　those　further　away.　In　this　study,　we　provide　an　improved　Mask　R-CNN　algorithm　which　is　also　effective　in　detecting　vehicles　for　a　low-angle　camera　perspective.　Firstly,　we　introduce　the　Selective　Kernel　Networks　(SKNet)　in　the　backbone　architectures.　Secondly,　we　build　a　path　with　clean　lateral　connections　from　the　low　level　to　the　top　ones　at　the　back　of　Feature　Pyramid　Networks　(FPN).　Thirdly,　we　replace　the　Non-Maximum　Suppression　(NMS)　with　the　Soft-NMS.　Compared　to　the　original　Mask　R-CNN,　the　improved　ones　have　better　performance,　particularly　for　a　low-angle　camera　perspective.