diff --git a/ruoyi-video/src/main/java/com/ruoyi/video/common/ModelManager.java b/ruoyi-video/src/main/java/com/ruoyi/video/common/ModelManager.java
index a9cc59a..54cc4cd 100644
--- a/ruoyi-video/src/main/java/com/ruoyi/video/common/ModelManager.java
+++ b/ruoyi-video/src/main/java/com/ruoyi/video/common/ModelManager.java
@@ -32,8 +32,8 @@ public final class ModelManager implements AutoCloseable {
             int rgb = palette[i % palette.length]; i++;
             int bgr = ((rgb & 0xFF) << 16) | (rgb & 0xFF00) | ((rgb >> 16) & 0xFF);
 
-            // 使用OpenVinoYoloDetector，但强制使用OpenCV后端
-            YoloDetector det = new OpenVinoYoloDetector(name, dir, w, h, "opencv", bgr);
+            // 使用OnnxYoloDetector替代OpenVinoYoloDetector
+            YoloDetector det = new OnnxYoloDetector(name, dir, w, h, backend, bgr);
             map.put(name, det);
         }
     }
@@ -45,4 +45,4 @@ public final class ModelManager implements AutoCloseable {
         map.values().forEach(d -> { try { d.close(); } catch(Exception ignored){} });
         map.clear();
     }
-}
+}
\ No newline at end of file
diff --git a/ruoyi-video/src/main/java/com/ruoyi/video/thread/detector/OnnxYoloDetector.java b/ruoyi-video/src/main/java/com/ruoyi/video/thread/detector/OnnxYoloDetector.java
new file mode 100644
index 0000000..44156ed
--- /dev/null
+++ b/ruoyi-video/src/main/java/com/ruoyi/video/thread/detector/OnnxYoloDetector.java
@@ -0,0 +1,249 @@
+package com.ruoyi.video.thread.detector;
+
+import com.ruoyi.video.domain.Detection;
+import org.bytedeco.javacpp.indexer.FloatRawIndexer;
+import org.bytedeco.opencv.opencv_core.*;
+import org.bytedeco.opencv.opencv_dnn.Net;
+
+import java.nio.file.*;
+import java.util.*;
+
+import static org.bytedeco.opencv.global.opencv_dnn.*;
+import static org.bytedeco.opencv.global.opencv_core.*;
+import static org.bytedeco.opencv.global.opencv_imgproc.*;
+
+public final class OnnxYoloDetector implements YoloDetector {
+    private final String modelName;
+    private final Net net;
+    private final Size input;
+    private final float confTh = 0.25f, nmsTh = 0.45f;
+    private final String[] classes;
+    private final int colorBGR;
+
+    public OnnxYoloDetector(String name, Path dir, int inW, int inH, String backend, int colorBGR) throws Exception {
+        this.modelName = name;
+        this.input = new Size(inW, inH);
+        this.colorBGR = colorBGR;
+
+        // 查找ONNX模型文件
+        String onnx = findModelFile(dir, ".onnx");
+        if (onnx == null) {
+            throw new Exception("找不到ONNX模型文件，请确保目录中存在 .onnx 文件: " + dir);
+        }
+
+        // 读取类别文件
+        Path clsPath = dir.resolve("classes.txt");
+        if (Files.exists(clsPath)) {
+            this.classes = Files.readAllLines(clsPath).stream().map(String::trim)
+                    .filter(s -> !s.isEmpty()).toArray(String[]::new);
+        } else {
+            this.classes = new String[0];
+        }
+
+        try {
+            // 加载ONNX模型
+            this.net = readNetFromONNX(onnx);
+            
+            // 设置OpenCV后端
+            net.setPreferableBackend(DNN_BACKEND_OPENCV);
+            net.setPreferableTarget(DNN_TARGET_CPU);
+            
+            System.out.println("ONNX模型加载成功: " + name + " (" + onnx + ")");
+            
+        } catch (Exception e) {
+            throw new Exception("模型加载失败: " + e.getMessage() + 
+                              "\n请确保ONNX模型文件格式正确", e);
+        }
+    }
+
+    /**
+     * 在目录中查找指定扩展名的模型文件
+     */
+    private String findModelFile(Path dir, String extension) {
+        try {
+            return Files.list(dir)
+                    .filter(path -> path.toString().toLowerCase().endsWith(extension.toLowerCase()))
+                    .map(Path::toString)
+                    .findFirst()
+                    .orElse(null);
+        } catch (Exception e) {
+            return null;
+        }
+    }
+
+    @Override public String name() { return modelName; }
+
+    @Override
+    public List<Detection> detect(Mat bgr) {
+        if (bgr == null || bgr.empty()) return Collections.emptyList();
+
+        // 统一成 BGR 3 通道，避免 blobFromImage 断言失败
+        if (bgr.channels() != 3) {
+            Mat tmp = new Mat();
+            if (bgr.channels() == 1) cvtColor(bgr, tmp, COLOR_GRAY2BGR);
+            else if (bgr.channels() == 4) cvtColor(bgr, tmp, COLOR_BGRA2BGR);
+            else bgr.copyTo(tmp);
+            bgr = tmp;
+        }
+
+        try (Mat blob = blobFromImage(bgr, 1.0/255.0, input, new Scalar(0.0), true, false, CV_32F)) {
+            net.setInput(blob);
+            // ===== 多输出兼容（Bytedeco 正确写法）=====
+            org.bytedeco.opencv.opencv_core.StringVector outNames = net.getUnconnectedOutLayersNames();
+            List<Mat> outs = new ArrayList<>();
+
+            if (outNames == null || outNames.size() == 0) {
+                // 只有一个默认输出
+                Mat out = net.forward();      // ← 直接返回 Mat
+                outs.add(out);
+            } else {
+                // 多输出：用 MatVector 承接
+                org.bytedeco.opencv.opencv_core.MatVector outBlobs =
+                        new org.bytedeco.opencv.opencv_core.MatVector(outNames.size());
+                net.forward(outBlobs, outNames);   // ← 正确的重载
+
+                for (long i = 0; i < outBlobs.size(); i++) {
+                    outs.add(outBlobs.get(i));
+                }
+            }
+
+            int fw = bgr.cols(), fh = bgr.rows();
+            List<Rect2d> boxes = new ArrayList<>();
+            List<Float> scores = new ArrayList<>();
+            List<Integer> classIds = new ArrayList<>();
+
+            for (Mat out : outs) {
+                parseYoloOutput(out, fw, fh, boxes, scores, classIds);
+            }
+            if (boxes.isEmpty()) return Collections.emptyList();
+
+            // 纯 Java NMS，避免 MatOf* / Vector API 兼容问题
+            List<Integer> keep = nmsIndices(boxes, scores, nmsTh);
+
+            List<Detection> result = new ArrayList<>(keep.size());
+            for (int k : keep) {
+                Rect2d r = boxes.get(k);
+                Rect rect = new Rect((int)r.x(), (int)r.y(), (int)r.width(), (int)r.height());
+                int cid = classIds.get(k);
+                String cname = (cid >= 0 && cid < classes.length) ? classes[cid] : ("cls"+cid);
+                result.add(new Detection("["+modelName+"] "+cname, scores.get(k), rect, colorBGR));
+            }
+            return result;
+        } catch (Throwable e) {
+            // 单帧失败不影响整体
+            return Collections.emptyList();
+        }
+    }
+
+    /** 解析 YOLO-IR 输出为 N×C（C>=6），并填充 boxes/scores/classIds。 */
+    private void parseYoloOutput(Mat out, int fw, int fh,
+                                 List<Rect2d> boxes, List<Float> scores, List<Integer> classIds) {
+        int dims = out.dims();
+        Mat m;
+
+        if (dims == 2) {
+            // NxC 或 CxN
+            if (out.cols() >= 6) {
+                m = out;
+            } else {
+                Mat tmp = new Mat();
+                transpose(out, tmp); // CxN -> NxC
+                m = tmp;
+            }
+        } else if (dims == 3) {
+            // [1,N,C] 或 [1,C,N]
+            if (out.size(2) >= 6) {
+                m = out.reshape(1, out.size(1)); // -> N×C
+            } else {
+                Mat squeezed = out.reshape(1, out.size(1)); // C×N
+                Mat tmp = new Mat();
+                transpose(squeezed, tmp);                   // -> N×C
+                m = tmp;
+            }
+        } else if (dims == 4) {
+            // [1,1,N,C] 或 [1,1,C,N]
+            int a = out.size(2), b = out.size(3);
+            if (b >= 6) {
+                m = out.reshape(1, a).clone(); // -> N×C
+            } else {
+                Mat cxn = out.reshape(1, b);   // C×N
+                Mat tmp = new Mat();
+                transpose(cxn, tmp);           // -> N×C
+                m = tmp.clone();
+            }
+        } else {
+            return; // 不支持的形状
+        }
+
+        int N = m.rows(), C = m.cols();
+        if (C < 6 || N <= 0) return;
+
+        FloatRawIndexer idx = m.createIndexer();
+        for (int i = 0; i < N; i++) {
+            float cx = idx.get(i,0), cy = idx.get(i,1), w = idx.get(i,2), h = idx.get(i,3);
+            float obj = idx.get(i,4);
+
+            int bestCls = -1; float bestScore = 0f;
+            for (int c = 5; c < C; c++) {
+                float p = idx.get(i,c);
+                if (p > bestScore) { bestScore = p; bestCls = c - 5; }
+            }
+            float conf = obj * bestScore;
+            if (conf < confTh) continue;
+
+            // 默认假设归一化中心点格式 (cx,cy,w,h)；若你的 IR 是 x1,y1,x2,y2，请把这里换算改掉
+            int bx = Math.max(0, Math.round(cx * fw - (w * fw) / 2f));
+            int by = Math.max(0, Math.round(cy * fh - (h * fh) / 2f));
+            int bw = Math.min(fw - bx, Math.round(w * fw));
+            int bh = Math.min(fh - by, Math.round(h * fh));
+            if (bw <= 0 || bh <= 0) continue;
+
+            boxes.add(new Rect2d(bx, by, bw, bh));
+            scores.add(conf);
+            classIds.add(bestCls);
+        }
+    }
+
+    /** 纯 Java NMS（IoU 抑制），返回保留的下标列表。 */
+    private List<Integer> nmsIndices(List<Rect2d> boxes, List<Float> scores, float nmsThreshold) {
+        List<Integer> order = new ArrayList<>(boxes.size());
+        for (int i = 0; i < boxes.size(); i++) order.add(i);
+        // 按分数降序
+        order.sort((i, j) -> Float.compare(scores.get(j), scores.get(i)));
+
+        List<Integer> keep = new ArrayList<>();
+        boolean[] removed = new boolean[boxes.size()];
+
+        for (int a = 0; a < order.size(); a++) {
+            int i = order.get(a);
+            if (removed[i]) continue;
+            keep.add(i);
+
+            Rect2d bi = boxes.get(i);
+            double areaI = bi.width() * bi.height();
+
+            for (int b = a + 1; b < order.size(); b++) {
+                int j = order.get(b);
+                if (removed[j]) continue;
+
+                Rect2d bj = boxes.get(j);
+                double areaJ = bj.width() * bj.height();
+
+                double xx1 = Math.max(bi.x(), bj.x());
+                double yy1 = Math.max(bi.y(), bj.y());
+                double xx2 = Math.min(bi.x() + bi.width(),  bj.x() + bj.width());
+                double yy2 = Math.min(bi.y() + bi.height(), bj.y() + bj.height());
+
+                double w = Math.max(0, xx2 - xx1);
+                double h = Math.max(0, yy2 - yy1);
+                double inter = w * h;
+                double iou = inter / (areaI + areaJ - inter + 1e-9);
+
+                if (iou > nmsThreshold) removed[j] = true;
+            }
+        }
+        return keep;
+    }
+
+    @Override public void close(){ net.close(); }
+} 
\ No newline at end of file
diff --git a/ruoyi-video/src/main/resources/libs/models/models.json b/ruoyi-video/src/main/resources/libs/models/models.json
index 14a7b7c..38e5aa4 100644
--- a/ruoyi-video/src/main/resources/libs/models/models.json
+++ b/ruoyi-video/src/main/resources/libs/models/models.json
@@ -1,4 +1,4 @@
 [
-  {"name":"smoke","path":"libs/models/smoke","size":[640,640],"backend":"OpenCV"},
-  {"name":"garbage","path":"libs/models/garbage","size":[640,640],"backend":"OpenCV"}
+  {"name":"smoke","path":"libs/models/smoke","size":[640,640],"backend":"opencv"},
+  {"name":"garbage","path":"libs/models/garbage","size":[640,640],"backend":"opencv"}
 ]