diff --git a/check.py b/check.py
new file mode 100644
index 0000000..21607a2
--- /dev/null
+++ b/check.py
@@ -0,0 +1,42 @@
+import cv2
+import os
+
+video_path = "./output_result.mp4"  # 替换成你上传的视频文件名
+
+if not os.path.exists(video_path):
+    print(f"错误：找不到文件 {video_path}")
+    exit()
+
+cap = cv2.VideoCapture(video_path)
+
+if not cap.isOpened():
+    print("错误：无法打开视频文件！")
+    exit()
+
+# 获取视频自带的物理帧率 (FPS)
+fps = cap.get(cv2.CAP_PROP_FPS)
+
+# 获取视频的总帧数
+total_frames = cap.get(cv2.CAP_PROP_FRAME_COUNT)
+
+# 获取视频的分辨率宽度和高度
+width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+# 计算视频总时长（秒）
+duration = total_frames / fps if fps > 0 else 0
+
+print("==========================================")
+print(f"视频文件: {os.path.basename(video_path)}")
+print(f"视频分辨率: {width} x {height}")
+print(f"视频原生帧率 (FPS): {fps:.2f} 帧/秒  <--- [核心原因看这里]")
+print(f"视频总帧数: {int(total_frames)} 帧")
+print(f"视频总时长: {duration:.2f} 秒")
+print("==========================================")
+
+if fps <= 25.5:
+    print("💡 结论：不出所料，你的视频本身就被录制/压制成了 25 帧！")
+    print("OpenCV 的 cap.read() 会卡在 25fps 的播放速度上等下一帧，导致 NPU 闲置。")
+    print("你想测试 NPU 的真实上限，需要用前面提到的『纯随机数矩阵压测脚本』。")
+
+cap.release()
\ No newline at end of file
diff --git a/infer_video.py b/infer_video.py
index 5173558..02d64dc 100644
--- a/infer_video.py
+++ b/infer_video.py
@@ -2,15 +2,53 @@ import os
 import cv2
 import numpy as np
 import time
+import threading
+import queue
+import socket
 from ais_bench.infer.interface import InferSession
 
-# ==================== 模拟测试硬编码配置 ====================
-MODEL_PATH = "./yolo26_8t.om"  # 你的 YOLO26 模型路径
-VIDEO_PATH = "./drone_test.mp4"  # 用于模拟相机的测试视频文件
+# ==================== 配置区域 ====================
+MODEL_PATH = "./om/yolo26n-balloon.om"  # YOLO26 模型路径
+INPUT_VIDEO = "./test_video.mp4"  # 输入的测试视频（首飞时可改成 0 代表机载 USB 相机）
 CONF_THRESHOLD = 0.25  # 置信度阈值
 
+# -------------- 无线图传网络配置 --------------
+GROUND_STATION_IP = "192.168.31.239"  # 地面站的局域网 IP
+UDP_PORT = 9999
+# ==================================================
+
+# 初始化一个图传专用的轻量级队列，最大长度为2。如果后台发得慢导致队列满了，主推理环会触发丢帧，绝对不卡死。
+tx_queue = queue.Queue(maxsize=2)
+
+
+def udp_stream_sender_thread():
+    """ 📡 后台图传专用线程 """
+    sock = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+    server_address = (GROUND_STATION_IP, UDP_PORT)
+    print(f"📡 图传后台线程已启动，正在持续向地面站 [{GROUND_STATION_IP}:{UDP_PORT}] 推流...")
+
+    while True:
+        try:
+            # 阻塞等待主线程丢进来的实时画框图像
+            frame = tx_queue.get()
+
+            # 1. 预处理：将图片缩放到 640x480（极大降低空中无线带宽压力和延迟）
+            small_frame = cv2.resize(frame, (640, 480))
+
+            # 2. 图像进行轻量级 JPEG 压缩（质量设为 60，画面清晰且计算速度极快）
+            ret, jpeg_img = cv2.imencode('.jpg', small_frame, [int(cv2.IMWRITE_JPEG_QUALITY), 60])
+
+            if ret:
+                data = np.array(jpeg_img).tobytes()
+                # 3. 确保单个 UDP 数据包不超过缓冲区上限（~65KB）
+                if len(data) < 65000:
+                    sock.sendto(data, server_address)
+
+            tx_queue.task_done()
+        except Exception:
+            # 即使无线网络抖动或地面站未开启，后台线程也静默承受，绝不连累无人机主系统
+            pass
 
-# ============================================================
 
 class YOLO26UAVInfer:
     def __init__(self, model_path, device_id=0):
@@ -20,100 +58,150 @@ class YOLO26UAVInfer:
         self.input_width = 640
         self.input_height = 640
 
-    def letterbox(self, img, new_shape=(640, 640), color=(114, 114, 114)):
+    def letterbox(self, img, new_shape=(640, 640), color=(114, 114, 114)) :
         shape = img.shape[:2]
         r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
         new_unpad = (int(round(shape[1] * r)), int(round(shape[0] * r)))
         dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]
-        left, top = dw // 2, dh // 2
-        right, bottom = dw - left, dh - top
+        dw /= 2
+        dh /= 2
         if shape[::-1] != new_unpad:
             img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
+        top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+        left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
         img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)
-        return img, r, (left, top)
+        return img, r, (dw, dh)
 
-    def preprocess_frame(self, img_bgr):
-        self.original_height, self.original_width = img_bgr.shape[:2]
-        img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
-        img_resized, self.ratio, (self.pad_left, self.pad_top) = self.letterbox(
-            img_rgb, (self.input_width, self.input_height)
-        )
-        img_input = (img_resized.astype(np.float32) / 255.0).astype(np.float32)
-        img_input = np.ascontiguousarray(np.transpose(img_input, (2, 0, 1))[np.newaxis, :, :, :])
-        return img_input
 
-    def infer(self, img_input):
-        return self.session.infer([img_input])
+def preprocess(self, img):
+    img_canvas, ratio, (dw, dh) = self.letterbox(img, (self.input_width, self.input_height))
+    img_in = img_canvas[:, :, ::-1].transpose(2, 0, 1)
+    img_in = np.ascontiguousarray(img_in).astype(np.float32) / 255.0
+    img_in = np.expand_dims(img_in, axis=0)
+    return img_in, ratio, (dw, dh)
 
-    def postprocess(self, infer_result, conf_threshold):
-        detections = []
-        output = infer_result[0] if isinstance(infer_result, (list, tuple)) else infer_result
-        if output.ndim == 3: output = output[0]
-        for det in output:
-            confidence = det[4]
-            if confidence < conf_threshold: continue
-            x1_scale, y1_scale, x2_scale, y2_scale = det[0], det[1], det[2], det[3]
-            cls_id = int(det[5])
-            x1 = max(0, min((x1_scale - self.pad_left) / self.ratio, self.original_width))
-            y1 = max(0, min((y1_scale - self.pad_top) / self.ratio, self.original_height))
-            x2 = max(0, min((x2_scale - self.pad_left) / self.ratio, self.original_width))
-            y2 = max(0, min((y2_scale - self.pad_top) / self.ratio, self.original_height))
-            detections.append([x1, y1, x2, y2, confidence, cls_id])
-        return detections
+
+def infer(self, input_tensor):
+    outputs = self.session.infer([input_tensor])
+    return outputs
+
+
+def postprocess(self, outputs, ratio, pad, orig_shape, conf_thres=0.25):
+    det_out = outputs[0][0]
+    valid_detections = []
+    for i in range(det_out.shape[0]):
+        row = det_out[i]
+        conf = row[4]
+        if conf < conf_thres:
+            continue
+        cls_id = int(row[5])
+        cx, cy, w, h = row[0], row[1], row[2], row[3]
+        x1 = cx - w / 2
+        y1 = cy - h / 2
+        x2 = cx + w / 2
+        y2 = cy + h / 2
+        dw, dh = pad
+        x1 = (x1 - dw) / ratio
+        y1 = (y1 - dh) / ratio
+        x2 = (x2 - dw) / ratio
+        y2 = (y2 - dh) / ratio
+        x1 = max(0, min(orig_shape[1], x1))
+        y1 = max(0, min(orig_shape[0], y1))
+        x2 = max(0, min(orig_shape[1], x2))
+        y2 = max(0, min(orig_shape[0], y2))
+        valid_detections.append([x1, y1, x2, y2, conf, cls_id])
+    return valid_detections
+
+
+def draw_boxes(img, detections):
+    img_copy = img.copy()
+    for det in detections:
+        x1, y1, x2, y2, conf, cls_id = det
+        x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
+        cv2.rectangle(img_copy, (x1, y1), (x2, y2), (0, 255, 0), 2)
+        label = f"Target: {conf:.2f}"
+        (label_w, label_h), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
+        cv2.rectangle(img_copy, (x1, y1 - label_h - 5), (x1 + label_w, y1), (0, 255, 0), -1)
+        cv2.putText(img_copy, label, (x1, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1)
+    return img_copy
 
 
 if __name__ == "__main__":
-    print("====== 启动香橙派无人机检测算法模拟测试 ======")
+    # 0. 提前拉起后台异步无线图传线程
+    stream_thread = threading.Thread(target=udp_stream_sender_thread, daemon=True)
+    stream_thread.start()
+
+    # 初始化推理会话
     yolo_infer = YOLO26UAVInfer(MODEL_PATH)
 
-    # 用视频文件模拟机载相机输入
-    cap = cv2.VideoCapture(VIDEO_PATH)
+    # 视频输入源
+    cap = cv2.VideoCapture(INPUT_VIDEO)
     if not cap.isOpened():
-        print(f"错误：无法打开模拟视频文件 {VIDEO_PATH}")
+        print(f"❌ 无法打开输入视频: {INPUT_VIDEO}")
         exit()
 
-    frame_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-    frame_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
-    img_center_x, img_center_y = frame_w / 2, frame_h / 2
-
-    print(f"成功加载视频，分辨率: {frame_w}x{frame_h} | 正在进行 NPU 循环推理...")
-    fps_start_time = time.time()
-    frame_count = 0
+    print("🚀 NPU 推理主循环已启动！")
 
     while True:
+        t_start = time.time()
         ret, frame = cap.read()
         if not ret:
-            print("视频播放结束或无法读取。")
+            print("视频播放结束或未捕获到帧。")
             break
 
-        frame_count += 1
+        orig_shape = frame.shape[:2]
+        img_center_x = orig_shape[1] / 2
+        img_center_y = orig_shape[0] / 2
 
-        # 1. 预处理与 NPU 推理
-        input_tensor = yolo_infer.preprocess_frame(frame)
-        raw_result = yolo_infer.infer(input_tensor)
-        detections = yolo_infer.postprocess(raw_result, conf_threshold=CONF_THRESHOLD)
+        # 1. 预处理
+        input_tensor, ratio, pad = yolo_infer.preprocess(frame)
 
-        # 2. 模拟数据下发接口（核心优化点）
+        # 2. 执行昇腾 NPU 硬件推理
+        outputs = yolo_infer.infer(input_tensor)
+
+        # 3. 后处理
+        detections = yolo_infer.postprocess(outputs, ratio, pad, orig_shape, conf_thres=CONF_THRESHOLD)
+
+        # 4. 执行你的核心控制逻辑（寻找最显著的目标并计算偏差）
         if len(detections) > 0:
-            best_det = max(detections, key=lambda x: x[4])  # 追踪置信度最高的目标
+            best_det = max(detections, key=lambda x: x[4])
             x1, y1, x2, y2, conf, cls_id = best_det
-            target_x, target_y = (x1 + x2) / 2, (y1 + y2) / 2
+            target_x = (x1 + x2) / 2
+            target_y = (y1 + y2) / 2
 
-            # 计算标准化像素偏差（-1.0 到 1.0）
             err_x = (target_x - img_center_x) / img_center_x
             err_y = (target_y - img_center_y) / img_center_y
 
-            # 模拟发送给飞控
-            print(
-                f"[Frame {frame_count:04d}] 检测到目标(类别:{cls_id}/置信度:{conf:.2f}) -> 模拟下发飞控控制量 err_x: {err_x:+.4f}, err_y: {err_y:+.4f}")
+            # 画一条从画面中心指向目标中心的蓝色对准线，直观展示控制走向
+            cv2.line(frame, (int(img_center_x), int(img_center_y)),
+                     (int(target_x), int(target_y)), (255, 0, 0), 2)
+
+            # 把计算好的、原本要发给飞控的误差值实时写在视频上
+            cv2.putText(frame, f"MAVLink Out -> Err_X: {err_x:+.3f} Err_Y: {err_y:+.3f}",
+                        (20, 70), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 0), 2)
         else:
-            print(f"[Frame {frame_count:04d}] 未检测到目标 -> 告知飞控保持悬停对准...")
+            cv2.putText(frame, "MAVLink Out -> Target LOST (HOVER)",
+                        (20, 70), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 255), 2)
 
-        # 3. 计算端到端真实 FPS（包含预处理、推理和后处理）
-        if frame_count % 30 == 0:
-            end_time = time.time()
-            avg_fps = 30 / (end_time - fps_start_time)
-            print(f"----------------> 当前香橙派端到端平均处理速度: {avg_fps:.1f} FPS")
-            fps_start_time = time.time()
+        # 5. 渲染基础边界框
+        frame = draw_boxes(frame, detections)
 
-    cap.release()
\ No newline at end of file
+        # 6. 计算单帧总耗时和纯计算 FPS
+        t_end = time.time()
+        fps = 1.0 / (t_end - t_start)
+        cv2.putText(frame, f"NPU Process-FPS: {fps:.1f}", (20, 40),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
+
+        # ==================== 异步丢帧推流 ====================
+        try:
+            # block=False 意味着如果无线网络太慢队列满了，程序不会原地死等，而是抛出异常并直接走下一次循环
+            tx_queue.put(frame, block=False)
+        except queue.Full:
+            pass  # 主动丢帧
+        # ===================================================================
+
+        # 打印调试信息到终端
+        print(f"当前帧推理完成 | 纯计算主环帧率: {fps:.1f} FPS")
+
+    cap.release()
+    print("香橙派端任务运行完毕。")
\ No newline at end of file
diff --git a/output_result.mp4 b/output_result.mp4
new file mode 100644
index 0000000..9e32a00
Binary files /dev/null and b/output_result.mp4 differ