diff --git a/bench_npu.py b/bench_npu.py
new file mode 100644
index 0000000..d0f7c59
--- /dev/null
+++ b/bench_npu.py
@@ -0,0 +1,149 @@
+#!/usr/bin/env python3
+"""
+昇腾 310B4 NPU 推理性能基准测试
+测试项目：
+1. 确认 NPU 推理（非 CPU fallback）
+2. 单次推理延迟
+3. 连续推理吞吐 (FPS)
+4. 多 worker 并发
+5. 不同分辨率影响
+"""
+import os
+import sys
+import time
+import socket
+import struct
+import json
+import base64
+import threading
+import numpy as np
+import cv2
+from concurrent.futures import ThreadPoolExecutor
+
+SOCK_PATH = "/tmp/edge-infer.sock"
+
+def make_test_image(w=640, h=640):
+    """生成随机测试图片"""
+    img = np.random.randint(0, 255, (h, w, 3), dtype=np.uint8)
+    _, buf = cv2.imencode('.jpg', img, [cv2.IMWRITE_JPEG_QUALITY, 90])
+    return buf.tobytes()
+
+def infer_one(jpeg_bytes):
+    """发送一次推理请求，返回响应时间(ms)"""
+    sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    sock.connect(SOCK_PATH)
+    
+    msg = {
+        "stream_id": 0,
+        "device_id": "bench",
+        "url": "bench://test",
+        "ts": 1234567890.0,
+        "jpeg_b64": base64.b64encode(jpeg_bytes).decode("utf-8")
+    }
+    
+    data = json.dumps(msg).encode("utf-8")
+    
+    t0 = time.time()
+    sock.sendall(struct.pack(">I", len(data)) + data)
+    hdr = sock.recv(4)
+    length = struct.unpack(">I", hdr)[0]
+    result = json.loads(sock.recv(length).decode("utf-8"))
+    elapsed = (time.time() - t0) * 1000  # ms
+    
+    sock.close()
+    return elapsed, result
+
+def test_single_infer():
+    """单次推理延迟"""
+    print("\n=== 单次推理延迟测试 ===")
+    jpeg = make_test_image()
+    times = []
+    for i in range(5):
+        t, _ = infer_one(jpeg)
+        times.append(t)
+        print(f"  第 {i+1} 次: {t:.1f} ms")
+    
+    # 跳过第一次预热
+    times = times[1:]
+    avg = sum(times) / len(times)
+    print(f"\n  平均延迟 (排除预热): {avg:.1f} ms")
+    print(f"  理论 FPS: {1000/avg:.1f}")
+    return avg
+
+def test_throughput():
+    """连续推理吞吐"""
+    print("\n=== 连续推理吞吐测试 ===")
+    jpeg = make_test_image()
+    count = 30
+    t0 = time.time()
+    
+    for i in range(count):
+        _, _ = infer_one(jpeg)
+        if (i+1) % 10 == 0:
+            elapsed = time.time() - t0
+            print(f"  {i+1}/{count} 完成, 累计: {elapsed:.1f}s, FPS: {(i+1)/elapsed:.1f}")
+    
+    total = time.time() - t0
+    fps = count / total
+    print(f"\n  总计 {count} 帧: {total:.2f}s")
+    print(f"  吞吐: {fps:.1f} FPS")
+    print(f"  每帧延迟: {total/count*1000:.1f} ms")
+    return fps
+
+def test_concurrent(workers=4):
+    """多 worker 并发推理"""
+    print(f"\n=== {workers} 路并发测试 ===")
+    jpeg = make_test_image()
+    results_per_worker = []
+    
+    def worker_task(worker_id, n_frames=20):
+        times = []
+        for _ in range(n_frames):
+            t, _ = infer_one(jpeg)
+            times.append(t)
+        return worker_id, times
+    
+    t0 = time.time()
+    with ThreadPoolExecutor(max_workers=workers) as pool:
+        futures = [pool.submit(worker_task, i, 15) for i in range(workers)]
+        for f in futures:
+            wid, times = f.result()
+            results_per_worker.append((wid, sum(times)/len(times), max(times), min(times)))
+    
+    total = time.time() - t0
+    total_frames = workers * 15
+    total_fps = total_frames / total
+    
+    for wid, avg, mx, mn in results_per_worker:
+        print(f"  Worker {wid}: avg={avg:.1f}ms, max={mx:.1f}ms, min={mn:.1f}ms")
+    
+    print(f"\n  并发 {workers} 路: 总计 {total_frames} 帧, {total:.2f}s")
+    print(f"  总吞吐: {total_fps:.1f} FPS")
+    print(f"  单路等效 FPS: {total_fps/workers:.1f}")
+    return total_fps
+
+def test_power():
+    """读取 NPU 功耗信息"""
+    import subprocess
+    try:
+        result = subprocess.run(['npu-smi', 'info'], capture_output=True, text=True, timeout=5)
+        print("\n=== NPU 状态 ===")
+        print(result.stdout)
+    except Exception as e:
+        print(f"无法读取 NPU 状态: {e}")
+
+if __name__ == "__main__":
+    print("=" * 60)
+    print("  昇腾 310B4 + ACL 原生推理 性能基准测试")
+    print("=" * 60)
+    
+    test_power()
+    test_single_infer()
+    test_throughput()
+    test_concurrent(2)
+    test_concurrent(4)
+    test_concurrent(6)
+    
+    print("\n" + "=" * 60)
+    print("测试完成")
+    print("=" * 60)
diff --git a/config/edge.yaml b/config/edge.yaml
new file mode 100644
index 0000000..784978e
--- /dev/null
+++ b/config/edge.yaml
@@ -0,0 +1,13 @@
+edge_id: edge-demo-001
+cloud_url: http://101.36.73.102:8004
+edge_token: ""
+
+# RTSP 流地址（留空使用演示模式）
+rtsp_urls: []
+
+infer_socket: /tmp/edge-infer.sock
+infer_fps: 5
+infer_workers: 3
+conf_threshold: 0.5
+dedup_window_sec: 30
+version: 1.0.0
diff --git a/debug_model.py b/debug_model.py
new file mode 100644
index 0000000..16b213a
--- /dev/null
+++ b/debug_model.py
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""调试: 检查模型原始输出"""
+import numpy as np
+import cv2
+import acl
+
+ACL_MEMCPY_HOST_TO_DEVICE = 1
+ACL_MEMCPY_DEVICE_TO_HOST = 2
+
+img = cv2.imread("/home/强光车灯误报.png")
+print("Image shape:", img.shape)
+
+acl.init()
+acl.rt.set_device(0)
+ctx, _ = acl.rt.create_context(0)
+
+model_id, _ = acl.mdl.load_from_file("/root/AI-tianyan/model/model.om")
+desc = acl.mdl.create_desc()
+acl.mdl.get_desc(desc, model_id)
+
+in_sz = acl.mdl.get_input_size_by_index(desc, 0)
+print("Input size:", in_sz, "bytes")
+
+# Preprocess
+h0, w0 = img.shape[:2]
+inp_h, inp_w = 640, 640
+scale = min(inp_h / h0, inp_w / w0)
+nh, nw = int(h0 * scale), int(w0 * scale)
+resized = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_LINEAR)
+canvas = np.full((inp_h, inp_w, 3), 114, dtype=np.uint8)
+pad_top = (inp_h - nh) // 2
+pad_left = (inp_w - nw) // 2
+canvas[pad_top:pad_top + nh, pad_left:pad_left + nw] = resized
+rgb = cv2.cvtColor(canvas, cv2.COLOR_BGR2RGB)
+blob = rgb.astype(np.float32) / 255.0
+blob = np.ascontiguousarray(blob.transpose(2, 0, 1)[np.newaxis])
+print("Blob shape:", blob.shape, "dtype:", blob.dtype)
+
+# Allocate device input
+in_ds = acl.mdl.create_dataset()
+host_buf = np.ascontiguousarray(blob)
+host_addr = acl.util.bytes_to_ptr(host_buf.tobytes())
+in_buf, _ = acl.rt.malloc(in_sz, 0)
+acl.rt.memcpy(in_buf, in_sz, host_addr, in_sz, ACL_MEMCPY_HOST_TO_DEVICE)
+acl.mdl.add_dataset_buffer(in_ds, acl.create_data_buffer(in_buf, in_sz))
+
+# Allocate output
+out_num = acl.mdl.get_num_outputs(desc)
+out_sizes = [acl.mdl.get_output_size_by_index(desc, i) for i in range(out_num)]
+out_ds = acl.mdl.create_dataset()
+out_bufs = []
+for i in range(out_num):
+    buf, _ = acl.rt.malloc(out_sizes[i], 0)
+    out_bufs.append(buf)
+    acl.mdl.add_dataset_buffer(out_ds, acl.create_data_buffer(buf, out_sizes[i]))
+    print(f"Output[{i}] size: {out_sizes[i]} bytes")
+
+# Execute
+ret = acl.mdl.execute(model_id, in_ds, out_ds)
+print("Execute ret:", ret)
+
+# Get output
+for i in range(out_num):
+    sz = out_sizes[i]
+    host = np.zeros(sz, dtype=np.uint8)
+    host_addr = acl.util.bytes_to_ptr(host.tobytes())
+    acl.rt.memcpy(host_addr, sz, out_bufs[i], sz, ACL_MEMCPY_DEVICE_TO_HOST)
+    
+    dims_out, _ = acl.mdl.get_output_dims(desc, i)
+    d = dims_out['dims']
+    print(f"\nOutput[{i}] dims: {d}")
+    print(f"  Size: {sz} bytes")
+    
+    # Try FP16
+    fp16 = host.view(np.float16)
+    print(f"  FP16 shape: {fp16.shape}")
+    
+    reshaped = fp16.astype(np.float32).reshape(d)
+    print(f"  Reshaped: {reshaped.shape}")
+    print(f"  Min: {reshaped.min():.4f}, Max: {reshaped.max():.4f}, Mean: {reshaped.mean():.4f}")
+    
+    # Check if it's all zeros or NaN
+    nan_count = np.isnan(reshaped).sum()
+    zero_count = (reshaped == 0).sum()
+    print(f"  NaN count: {nan_count}, Zero count: {zero_count}/{reshaped.size}")
+    
+    # For YOLOv8 output [1, 84, 8400], check class scores
+    # boxes: reshaped[:, :4, :]
+    # scores: reshaped[:, 4:, :]
+    scores = reshaped[0, 4:, :]
+    max_scores = scores.max(axis=0)  # max class score per anchor
+    print(f"\n  Max class scores per anchor:")
+    print(f"    Min: {max_scores.min():.4f}, Max: {max_scores.max():.4f}, Mean: {max_scores.mean():.4f}")
+    
+    # Count anchors with score > 0.1
+    high_conf = (max_scores > 0.1).sum()
+    print(f"    Anchors with conf > 0.1: {high_conf}")
+    print(f"    Anchors with conf > 0.01: {(max_scores > 0.01).sum()}")
+    
+    # Print top 5
+    top5_idx = np.argsort(max_scores)[-5:][::-1]
+    for idx in top5_idx:
+        best_cls = scores[:, idx].argmax()
+        print(f"    Anchor {idx}: class={best_cls} (score={max_scores[idx]:.4f})")
+        box = reshaped[0, :4, idx]
+        print(f"      box: {box}")
+
+# Cleanup
+for buf in out_bufs:
+    acl.rt.free(buf)
+acl.rt.free(in_buf)
+acl.mdl.destroy_dataset(in_ds)
+acl.mdl.destroy_dataset(out_ds)
+acl.mdl.unload(model_id)
+acl.mdl.destroy_desc(desc)
+acl.rt.destroy_context(ctx)
+acl.rt.reset_device(0)
+acl.finalize()
+print("\nDone")
diff --git a/detect_one.py b/detect_one.py
new file mode 100644
index 0000000..81d1e85
--- /dev/null
+++ b/detect_one.py
@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""检测单张图片 - 通过 Unix socket 发送到推理服务"""
+import sys
+import socket
+import struct
+import json
+import base64
+import cv2
+
+SOCK_PATH = "/tmp/edge-infer.sock"
+
+def detect_image(image_path):
+    img = cv2.imread(image_path)
+    if img is None:
+        print("无法读取图片:", image_path)
+        return
+    
+    # 编码为 JPEG
+    _, buf = cv2.imencode('.jpg', img, [cv2.IMWRITE_JPEG_QUALITY, 90])
+    jpeg_bytes = buf.tobytes()
+    
+    # 连接推理服务
+    sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    sock.connect(SOCK_PATH)
+    
+    # 构造请求
+    msg = {
+        "stream_id": 0,
+        "device_id": "cli-test",
+        "url": "file://" + image_path,
+        "ts": 1234567890.0,
+        "jpeg_b64": base64.b64encode(jpeg_bytes).decode("utf-8")
+    }
+    
+    # 发送
+    data = json.dumps(msg).encode("utf-8")
+    sock.sendall(struct.pack(">I", len(data)) + data)
+    
+    # 接收结果
+    hdr = sock.recv(4)
+    if not hdr:
+        print("未收到响应")
+        sock.close()
+        return
+    
+    length = struct.unpack(">I", hdr)[0]
+    result = json.loads(sock.recv(length).decode("utf-8"))
+    sock.close()
+    
+    # 输出结果
+    dets = result.get("detections", [])
+    print(f"\n图片: {image_path}")
+    print(f"尺寸: {img.shape[1]}x{img.shape[0]}")
+    print(f"检测到 {len(dets)} 个目标\n")
+    print(f"{'类别':<20} {'置信度':<10} {'边界框'}")
+    print("-" * 60)
+    
+    for d in dets:
+        bbox = d["bbox"]
+        print(f'{d["class"]:<20} {d["conf"]:<10.3f} [{bbox[0]:.0f}, {bbox[1]:.0f}, {bbox[2]:.0f}, {bbox[3]:.0f}]')
+    
+    # 保存带标注的图片
+    if dets:
+        for d in dets:
+            x1, y1, x2, y2 = [int(x) for x in d["bbox"]]
+            cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)
+            label = f'{d["class"]} {d["conf"]:.2f}'
+            cv2.putText(img, label, (x1, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
+        
+        out_path = image_path.rsplit(".", 1)[0] + "_result.jpg"
+        cv2.imwrite(out_path, img)
+        print(f"\n已保存标注图片: {out_path}")
+
+if __name__ == "__main__":
+    path = sys.argv[1] if len(sys.argv) > 1 else "/home/强光车灯误报.png"
+    detect_image(path)
diff --git a/go.sum b/go.sum
new file mode 100644
index 0000000..33d593e
--- /dev/null
+++ b/go.sum
@@ -0,0 +1,10 @@
+github.com/kr/pretty v0.2.1 h1:Fmg33tUaq4/8ym9TJN1x7sLJnHVwhP33CNkpYV/7rwI=
+github.com/kr/pretty v0.2.1/go.mod h1:ipq/a2n7PKx3OHsz4KJII5eveXtPO4qwEXGdVfWzfnI=
+github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
+github.com/kr/text v0.1.0 h1:45sCR5RtlFHMR4UwH9sdQ5TC8v0qDQCHnXt+kaKSTVE=
+github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
+gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
diff --git a/model/model.om b/model/model.om
new file mode 100644
index 0000000..90a2456
Binary files /dev/null and b/model/model.om differ
diff --git a/model/names.txt b/model/names.txt
new file mode 100644
index 0000000..941cb4e
--- /dev/null
+++ b/model/names.txt
@@ -0,0 +1,80 @@
+person
+bicycle
+car
+motorcycle
+airplane
+bus
+train
+truck
+boat
+traffic light
+fire hydrant
+stop sign
+parking meter
+bench
+bird
+cat
+dog
+horse
+sheep
+cow
+elephant
+bear
+zebra
+giraffe
+backpack
+umbrella
+handbag
+tie
+suitcase
+frisbee
+skis
+snowboard
+sports ball
+kite
+baseball bat
+baseball glove
+skateboard
+surfboard
+tennis racket
+bottle
+wine glass
+cup
+fork
+knife
+spoon
+bowl
+banana
+apple
+sandwich
+orange
+broccoli
+carrot
+hot dog
+pizza
+donut
+cake
+chair
+couch
+potted plant
+bed
+dining table
+toilet
+tv
+laptop
+mouse
+remote
+keyboard
+cell phone
+microwave
+oven
+toaster
+sink
+refrigerator
+book
+clock
+vase
+scissors
+teddy bear
+hair drier
+toothbrush
diff --git a/test_infer.py b/test_infer.py
new file mode 100644
index 0000000..084e183
--- /dev/null
+++ b/test_infer.py
@@ -0,0 +1,79 @@
+#!/usr/bin/env python3
+"""测试推理服务 - 发送一张测试图片"""
+import socket
+import struct
+import json
+import base64
+import numpy as np
+import cv2
+import os
+
+SOCK_PATH = "/tmp/edge-infer.sock"
+
+def create_test_image():
+    """创建一张 640x480 的测试图片，画几个几何图形"""
+    img = np.zeros((480, 640, 3), dtype=np.uint8)
+    # 画一些简单的形状模拟检测目标
+    cv2.rectangle(img, (50, 50), (200, 200), (255, 255, 255), -1)
+    cv2.circle(img, (400, 300), 80, (255, 255, 255), -1)
+    # 编码为 JPEG
+    _, buf = cv2.imencode('.jpg', img)
+    return buf.tobytes()
+
+def send_msg(conn, payload):
+    data = json.dumps(payload).encode("utf-8")
+    conn.sendall(struct.pack(">I", len(data)) + data)
+
+def recv_msg(conn):
+    hdr = conn.recv(4)
+    if not hdr:
+        return None
+    length = struct.unpack(">I", hdr)[0]
+    data = b""
+    while len(data) < length:
+        chunk = conn.recv(length - len(data))
+        if not chunk:
+            return None
+        data += chunk
+    return json.loads(data.decode("utf-8"))
+
+def main():
+    if not os.path.exists(SOCK_PATH):
+        print(f"错误: socket {SOCK_PATH} 不存在")
+        return
+    
+    print("连接推理服务...")
+    conn = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    conn.connect(SOCK_PATH)
+    
+    # 创建测试图片
+    jpeg_bytes = create_test_image()
+    msg = {
+        "stream_id": 0,
+        "device_id": "test-001",
+        "url": "test://local",
+        "ts": 1234567890.0,
+        "jpeg_b64": base64.b64encode(jpeg_bytes).decode("utf-8")
+    }
+    
+    print("发送测试图片...")
+    send_msg(conn, msg)
+    
+    print("等待推理结果...")
+    result = recv_msg(conn)
+    
+    if result:
+        print(f"\n推理结果:")
+        print(f"  stream_id: {result['stream_id']}")
+        print(f"  device_id: {result['device_id']}")
+        print(f"  detections: {len(result['detections'])} 个目标")
+        for d in result['detections']:
+            print(f"    - {d['class']}: {d['conf']:.3f} bbox={d['bbox']}")
+    else:
+        print("未收到结果")
+    
+    conn.close()
+    print("\n测试完成!")
+
+if __name__ == "__main__":
+    main()
diff --git a/yolov8n.om b/yolov8n.om
new file mode 100644
index 0000000..90a2456
Binary files /dev/null and b/yolov8n.om differ