Fix NV12 crash issue when recreate camera object

(new structure) does not work

modules/ANSODEngine/CMakeLists.txt

@@ -14,6 +14,7 @@ set(ANSOD_SOURCES
     ANSFaceDetectorEngine.cpp
     ANSFaceRecognizerEngine.cpp
     ANSGpuFrameRegistry.cpp
+    GpuNV12SlotPool.cpp
     ANSONNXCL.cpp
     ANSONNXOBB.cpp
     ANSONNXPOSE.cpp

modules/ANSODEngine/GpuNV12SlotPool.cpp

@@ -0,0 +1,23 @@
+// GpuNV12SlotPool.cpp — Cross-DLL singleton resolver for ANSODEngine.dll.
+//
+// Finds the canonical GpuNV12SlotPool instance exported by ANSCV.dll
+// via GetProcAddress.  No link dependency on ANSCV.lib needed.
+
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include "GpuNV12SlotPool.h"
+
+GpuNV12SlotPool* GpuNV12SlotPool::resolveProcessWide() {
+    // ANSCV.dll is always loaded before inference starts (it provides RTSP).
+    HMODULE hMod = GetModuleHandleA("ANSCV.dll");
+    if (hMod) {
+        typedef GpuNV12SlotPool* (*GetInstanceFn)();
+        auto fn = reinterpret_cast<GetInstanceFn>(
+            GetProcAddress(hMod, "GpuNV12SlotPool_GetInstance"));
+        if (fn) return fn();
+    }
+    // Fallback: local instance (unit tests without ANSCV.dll).
+    static GpuNV12SlotPool local;
+    return &local;
+}

modules/ANSODEngine/NV12PreprocessHelper.cpp

@@ -276,6 +276,7 @@ namespace ANSCENTER {
         const bool useZeroCopy = isCudaDevice && gpuMatch;
 
         // --- Debug: log pointer state before reading ---
+#if defined(ANSCORE_GPU_DEBUG) && ANSCORE_GPU_DEBUG
         {
             char _nv12_dbg[512];
             snprintf(_nv12_dbg, sizeof(_nv12_dbg),
@@ -294,6 +295,7 @@ namespace ANSCENTER {
 #endif
             fprintf(stderr, "%s", _nv12_dbg);
         }
+#endif
 
         // Effective plane pointers — for zero-copy, use CUDA device ptrs;
         // for CPU upload, use the CPU snapshot buffers.
@@ -362,7 +364,7 @@ namespace ANSCENTER {
         cv::cuda::GpuMat gpuY, gpuUV;
 
         if (useZeroCopy) {
-            // CUDA zero-copy: wrap NVDEC device pointers directly
+            // CUDA zero-copy: wrap pool buffer device pointers directly
             gpuY  = cv::cuda::GpuMat(frameH,     frameW, CV_8UC1,
                         effYPlane,  static_cast<size_t>(effYLinesize));
             gpuUV = cv::cuda::GpuMat(frameH / 2, frameW, CV_8UC1,
@@ -455,6 +457,7 @@ namespace ANSCENTER {
         gpuResized.create(inputH, inputW, CV_8UC3);
 
         cudaStream_t rawStream = cv::cuda::StreamAccessor::getStream(stream);
+#if defined(ANSCORE_GPU_DEBUG) && ANSCORE_GPU_DEBUG
         {
             char _nv12_dbg2[256];
             snprintf(_nv12_dbg2, sizeof(_nv12_dbg2),
@@ -467,6 +470,7 @@ namespace ANSCENTER {
 #endif
             fprintf(stderr, "%s", _nv12_dbg2);
         }
+#endif
         launcher(gpuY, gpuUV, gpuResized, frameW, frameH, inputW, inputH, rawStream);
 
         stream.waitForCompletion();