// GrooveOccupancyDetector.cpp #include "GrooveOccupancyDetector.h" #include #include #include #include #include #include #include struct GrooveOccupancyDetector::OrtImpl { std::unique_ptr env; Ort::SessionOptions options; std::unique_ptr session; Ort::AllocatorWithDefaultOptions allocator; std::string inputName; std::string outputName; }; GrooveOccupancyDetector::GrooveOccupancyDetector() = default; GrooveOccupancyDetector::~GrooveOccupancyDetector() = default; void GrooveOccupancyDetector::setError(const QString& error) { m_lastError = error; qWarning() << "[GrooveOccupancyDetector]" << error; } bool GrooveOccupancyDetector::loadModel(const QString& onnxPath) { m_loaded = false; m_modelPath.clear(); m_lastError.clear(); m_ort.reset(); try { auto impl = std::make_unique(); impl->env = std::make_unique(ORT_LOGGING_LEVEL_WARNING, "GrooveOccupancyDetector"); impl->options.SetIntraOpNumThreads(4); impl->options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL); #ifdef _WIN32 const std::wstring widePath = onnxPath.toStdWString(); impl->session = std::make_unique(*impl->env, widePath.c_str(), impl->options); #else impl->session = std::make_unique(*impl->env, onnxPath.toUtf8().constData(), impl->options); #endif if (impl->session->GetInputCount() != 1 || impl->session->GetOutputCount() != 1) { setError(QStringLiteral("凹槽模型必须只有一个输入和一个输出。")); return false; } auto inputName = impl->session->GetInputNameAllocated(0, impl->allocator); auto outputName = impl->session->GetOutputNameAllocated(0, impl->allocator); impl->inputName = inputName.get(); impl->outputName = outputName.get(); const auto inputShape = impl->session->GetInputTypeInfo(0) .GetTensorTypeAndShapeInfo().GetShape(); if (inputShape.size() != 4 || inputShape[1] != 3 || inputShape[2] <= 0 || inputShape[3] <= 0) { setError(QStringLiteral("凹槽模型输入不是固定的 [1,3,H,W] RGB 张量。")); return false; } m_inputH = static_cast(inputShape[2]); m_inputW = static_cast(inputShape[3]); const auto outputShape = impl->session->GetOutputTypeInfo(0) .GetTensorTypeAndShapeInfo().GetShape(); if (outputShape.size() != 3 || outputShape[0] != 1 || outputShape[1] != 6 || outputShape[2] <= 0) { setError(QStringLiteral( "凹槽模型输出必须是 [1,6,N](xywh + empty/occupied 两类)。")); return false; } m_ort = std::move(impl); m_modelPath = onnxPath; m_loaded = true; qDebug() << "[GrooveOccupancyDetector] model loaded:" << onnxPath << "input:" << m_inputW << "x" << m_inputH; return true; } catch (const Ort::Exception& exception) { setError(QStringLiteral("ONNX Runtime 加载失败:%1").arg(exception.what())); } catch (const std::exception& exception) { setError(QStringLiteral("凹槽模型加载失败:%1").arg(exception.what())); } return false; } std::vector GrooveOccupancyDetector::preprocess(const QImage& image) const { const QImage rgb = image.convertToFormat(QImage::Format_RGB888); // scaled() can promote the result to RGB32. Convert back before reading // three bytes per pixel, so the CHW tensor is never byte-shifted. const QImage resized = rgb.scaled(m_inputW, m_inputH, Qt::IgnoreAspectRatio, Qt::SmoothTransformation) .convertToFormat(QImage::Format_RGB888); const size_t planeSize = static_cast(m_inputW) * m_inputH; std::vector tensor(planeSize * 3); for (int y = 0; y < m_inputH; ++y) { const uchar* line = resized.constScanLine(y); for (int x = 0; x < m_inputW; ++x) { const size_t index = static_cast(y) * m_inputW + x; tensor[index] = line[x * 3] / 255.f; tensor[planeSize + index] = line[x * 3 + 1] / 255.f; tensor[planeSize * 2 + index] = line[x * 3 + 2] / 255.f; } } return tensor; } QVector GrooveOccupancyDetector::nms( QVector candidates, float iouThreshold) const { std::sort(candidates.begin(), candidates.end(), [](const GrooveDetectionResult& left, const GrooveDetectionResult& right) { return left.confidence > right.confidence; }); QVector kept; QVector suppressed(candidates.size(), false); for (int i = 0; i < candidates.size(); ++i) { if (suppressed[i]) continue; kept.append(candidates[i]); const QRectF& a = candidates[i].box; for (int j = i + 1; j < candidates.size(); ++j) { if (suppressed[j] || candidates[i].classId != candidates[j].classId) continue; const QRectF& b = candidates[j].box; const float x1 = std::max(a.left(), b.left()); const float y1 = std::max(a.top(), b.top()); const float x2 = std::min(a.right(), b.right()); const float y2 = std::min(a.bottom(), b.bottom()); const float intersection = std::max(0.f, x2 - x1) * std::max(0.f, y2 - y1); const float unionArea = a.width() * a.height() + b.width() * b.height() - intersection; if (unionArea > 0.f && intersection / unionArea > iouThreshold) suppressed[j] = true; } } return kept; } QVector GrooveOccupancyDetector::postprocess( const float* output, int numAnchors, int numFields, int originalWidth, int originalHeight, float confidenceThreshold, float nmsThreshold) const { QVector candidates; candidates.reserve(128); const float scaleX = static_cast(originalWidth) / m_inputW; const float scaleY = static_cast(originalHeight) / m_inputH; for (int anchor = 0; anchor < numAnchors; ++anchor) { const float emptyScore = output[4 * numAnchors + anchor]; const float occupiedScore = output[5 * numAnchors + anchor]; const int classId = occupiedScore > emptyScore ? 1 : 0; const float score = classId == 0 ? emptyScore : occupiedScore; if (!std::isfinite(score) || score < confidenceThreshold) continue; const float centerX = output[anchor]; const float centerY = output[numAnchors + anchor]; const float width = output[2 * numAnchors + anchor]; const float height = output[3 * numAnchors + anchor]; if (!(width > 0.f && height > 0.f)) continue; const float x1 = qBound(0.f, (centerX - width * .5f) * scaleX, static_cast(originalWidth)); const float y1 = qBound(0.f, (centerY - height * .5f) * scaleY, static_cast(originalHeight)); const float x2 = qBound(0.f, (centerX + width * .5f) * scaleX, static_cast(originalWidth)); const float y2 = qBound(0.f, (centerY + height * .5f) * scaleY, static_cast(originalHeight)); if (x2 <= x1 || y2 <= y1) continue; candidates.append({ QRectF(x1, y1, x2 - x1, y2 - y1), classId, score }); } return nms(std::move(candidates), nmsThreshold); } QVector GrooveOccupancyDetector::detect(const QImage& image, float confidenceThreshold, float nmsThreshold) { if (!m_loaded || !m_ort || !m_ort->session) { setError(QStringLiteral("凹槽模型尚未加载。")); return {}; } if (image.isNull()) { setError(QStringLiteral("凹槽检测输入图像为空。")); return {}; } try { std::vector input = preprocess(image); const std::vector inputShape = { 1, 3, m_inputH, m_inputW }; const Ort::MemoryInfo memory = Ort::MemoryInfo::CreateCpu( OrtArenaAllocator, OrtMemTypeDefault); Ort::Value inputTensor = Ort::Value::CreateTensor(memory, input.data(), input.size(), inputShape.data(), inputShape.size()); const char* inputName = m_ort->inputName.c_str(); const char* outputName = m_ort->outputName.c_str(); auto outputs = m_ort->session->Run(Ort::RunOptions{ nullptr }, &inputName, &inputTensor, 1, &outputName, 1); if (outputs.empty()) { setError(QStringLiteral("凹槽模型没有返回输出张量。")); return {}; } const auto shape = outputs[0].GetTensorTypeAndShapeInfo().GetShape(); if (shape.size() != 3 || shape[0] != 1 || shape[1] != 6 || shape[2] <= 0) { setError(QStringLiteral("凹槽模型推理输出形状异常。")); return {}; } return postprocess(outputs[0].GetTensorData(), static_cast(shape[2]), static_cast(shape[1]), image.width(), image.height(), confidenceThreshold, nmsThreshold); } catch (const Ort::Exception& exception) { setError(QStringLiteral("凹槽 ONNX 推理失败:%1").arg(exception.what())); } catch (const std::exception& exception) { setError(QStringLiteral("凹槽检测失败:%1").arg(exception.what())); } return {}; }