diff --git a/modules/core/include/opencv2/core/cuda.hpp b/modules/core/include/opencv2/core/cuda.hpp
index f3da470899..f05d51dc9a 100644
--- a/modules/core/include/opencv2/core/cuda.hpp
+++ b/modules/core/include/opencv2/core/cuda.hpp
@@ -344,6 +344,8 @@ class CV_EXPORTS HostMem
 public:
     enum AllocType { PAGE_LOCKED = 1, SHARED = 2, WRITE_COMBINED = 4 };
 
+    static MatAllocator* getAllocator(AllocType alloc_type = PAGE_LOCKED);
+
     explicit HostMem(AllocType alloc_type = PAGE_LOCKED);
 
     HostMem(const HostMem& m);
diff --git a/modules/core/src/cuda_host_mem.cpp b/modules/core/src/cuda_host_mem.cpp
index dafa4f1621..2ad733b675 100644
--- a/modules/core/src/cuda_host_mem.cpp
+++ b/modules/core/src/cuda_host_mem.cpp
@@ -42,10 +42,124 @@
 //M*/
 
 #include "precomp.hpp"
+#include <map>
 
 using namespace cv;
 using namespace cv::cuda;
 
+#ifdef HAVE_CUDA
+
+namespace {
+
+class HostMemAllocator : public MatAllocator
+{
+public:
+    explicit HostMemAllocator(unsigned int flags) : flags_(flags)
+    {
+    }
+
+    UMatData* allocate(int dims, const int* sizes, int type,
+                       void* data0, size_t* step,
+                       int /*flags*/, UMatUsageFlags /*usageFlags*/) const
+    {
+        size_t total = CV_ELEM_SIZE(type);
+        for (int i = dims-1; i >= 0; i--)
+        {
+            if (step)
+            {
+                if (data0 && step[i] != CV_AUTOSTEP)
+                {
+                    CV_Assert(total <= step[i]);
+                    total = step[i];
+                }
+                else
+                {
+                    step[i] = total;
+                }
+            }
+
+            total *= sizes[i];
+        }
+
+        UMatData* u = new UMatData(this);
+        u->size = total;
+
+        if (data0)
+        {
+            u->data = u->origdata = static_cast<uchar*>(data0);
+            u->flags |= UMatData::USER_ALLOCATED;
+        }
+        else
+        {
+            void* ptr = 0;
+            cudaSafeCall( cudaHostAlloc(&ptr, total, flags_) );
+
+            u->data = u->origdata = static_cast<uchar*>(ptr);
+        }
+
+        return u;
+    }
+
+    bool allocate(UMatData* u, int /*accessFlags*/, UMatUsageFlags /*usageFlags*/) const
+    {
+        return (u != NULL);
+    }
+
+    void deallocate(UMatData* u) const
+    {
+        CV_Assert(u->urefcount >= 0);
+        CV_Assert(u->refcount >= 0);
+
+        if (u && u->refcount == 0)
+        {
+            if ( !(u->flags & UMatData::USER_ALLOCATED) )
+            {
+                cudaFreeHost(u->origdata);
+                u->origdata = 0;
+            }
+
+            delete u;
+        }
+    }
+
+private:
+    unsigned int flags_;
+};
+
+} // namespace
+
+#endif
+
+MatAllocator* cv::cuda::HostMem::getAllocator(AllocType alloc_type)
+{
+#ifndef HAVE_CUDA
+    (void) alloc_type;
+    throw_no_cuda();
+    return NULL;
+#else
+    static std::map<unsigned int, Ptr<MatAllocator> > allocators;
+
+    unsigned int flag = cudaHostAllocDefault;
+
+    switch (alloc_type)
+    {
+    case PAGE_LOCKED:    flag = cudaHostAllocDefault; break;
+    case SHARED:         flag = cudaHostAllocMapped;  break;
+    case WRITE_COMBINED: flag = cudaHostAllocWriteCombined; break;
+    default:             CV_Error(cv::Error::StsBadFlag, "Invalid alloc type");
+    }
+
+    Ptr<MatAllocator>& a = allocators[flag];
+
+    if (a.empty())
+    {
+        a = makePtr<HostMemAllocator>(flag);
+    }
+
+    return a.get();
+#endif
+}
+
 #ifdef HAVE_CUDA
 namespace
 {
diff --git a/modules/core/test/cuda/test_stream.cpp b/modules/core/test/cuda/test_stream.cpp
index bf3316eba3..a0e451a62a 100644
--- a/modules/core/test/cuda/test_stream.cpp
+++ b/modules/core/test/cuda/test_stream.cpp
@@ -129,6 +129,25 @@ CUDA_TEST_P(Async, Convert)
     stream.waitForCompletion();
 }
 
+CUDA_TEST_P(Async, HostMemAllocator)
+{
+    cv::cuda::Stream stream;
+
+    cv::Mat h_dst;
+    h_dst.allocator = cv::cuda::HostMem::getAllocator();
+
+    d_src.upload(src, stream);
+    d_src.convertTo(d_dst, CV_32S, stream);
+    d_dst.download(h_dst, stream);
+
+    stream.waitForCompletion();
+
+    cv::Mat dst_gold;
+    src.createMatHeader().convertTo(dst_gold, CV_32S);
+
+    ASSERT_MAT_NEAR(dst_gold, h_dst, 0);
+}
+
 INSTANTIATE_TEST_CASE_P(CUDA_Stream, Async, ALL_DEVICES);
 
 #endif // HAVE_CUDA