1.程序中ffmpeg启用硬件加速\n2.修改yolov8rknn以适配main程序

config.yaml

@@ -1,5 +1,6 @@
 
 engine_path: 'build_0826/'
+rknn_path: '/home/orangepi/Desktop/kangda_robotic_dog/yolov8_20250820.rknn'
 
 video_config:
 
@@ -124,7 +125,8 @@ dataConfig:
 #  putMessageUrl: 'http://192.168.220.202/api/edge/edgecallmanages/vi-alarm/v1'
 
 compreface_service:
-  domain: 'http://localhost'
+  # domain: 'http://localhost'
+  domain: 'http://10.0.0.202'
   port: '8000'
   api_key: 'ce04b456-88df-4df8-a7a2-69849111916f'
   # api_key: 'ab77978a-cc2b-4fa0-8959-6294e856721a'

main.py

@@ -89,20 +89,26 @@ m3u8_path = configData['video_config']['m3u8_path']
 save_path = configData['video_config']['save_path']
 vod_path = configData['video_config']['v1_path']
 people_save_path = configData['video_config']['people_save_path']
-vod_channelNo = configData['video_config']['v1_channelNo']
-testclasses = configData['video_config']['v1_testclasses']
+vod_channelNo = configData['video_config']['v0_channelNo']
+testclasses = configData['video_config']['v0_testclasses']
 engine_path = configData['engine_path']
+rknn_path = configData['rknn_path']
+
 
 # FFmpeg命令
 command_mid = [
     'ffmpeg',
+    '-hwaccel', 'rkmpp',  # 启用 RKMPP 硬件加速解码（硬件解码）
     '-i', '-',  # 从标准输入读取视频帧
-    '-c:v', 'libx264',  # 使用 H.264 编码
+    '-vf', 'scale_rkrga=1280:720',  # 使用 RGA 硬件加速缩放（零拷贝）
+    '-c:v', 'h264_rkmpp', # 使用 RKMPP 硬件加速 H.264 编码（代替 libx264）
     '-b:v', '500k',  # 设置视频比特率
-    '-preset', 'superfast',  # 编码速度
-    '-tune', 'zerolatency',  # 低延迟
-    '-crf', '23',  # 使用 CRF 模式来控制视频质量
-    '-s', '1280x720',  # 设置分辨率
+    '-g', '50', # GOP 大小（可按需要调节）, 硬件解码时需要
+    '-bf', '0',    # 禁用 B 帧以减少延迟, 硬件解码时需要
+    # '-preset', 'superfast',  # 编码速度, cpu编码时需要
+    # '-tune', 'zerolatency',  # 低延迟, cpu编码时需要
+    # '-crf', '23',  # 使用 CRF 模式来控制视频质量, cpu编码时需要
+    # '-s', '1280x720',  # 设置分辨率, cpu编码时需要
     '-an',  # 禁用音频
     '-loglevel', 'error',
     '-hls_time', '4',
@@ -845,7 +851,7 @@ class inferThread(threading.Thread):
                     # logger.debug(f"获取token: {token}")
 
                     try:
-                        batch_image_raw, r_list, box_list = self.yolov8_wrapper.infer(frame)
+                        batch_image_raw, r_list, box_list = self.yolov8_wrapper.infer(frame, CONF_THRESH, IOU_THRESHOLD)
                     except Exception as e:
                         logger.error(f"目标检测时发生异常: {e}")
                         # 即使检测失败，也要将帧推送到ffmpeg
@@ -1034,8 +1040,8 @@ if __name__ == "__main__":
     os.makedirs(m3u8_path + ip, exist_ok=True)
     
     # load custom plugin and engine
-    PLUGIN_LIBRARY = f"{engine_path}libmyplugins.so"
-    engine_file_path = f"{engine_path}best.engine"
+    # PLUGIN_LIBRARY = f"{engine_path}libmyplugins.so"
+    # engine_file_path = f"{engine_path}best.engine"
 
     # 执行python代码命令行 参数判断操作（不做处理）
     if len(sys.argv) > 1:
@@ -1045,19 +1051,19 @@ if __name__ == "__main__":
         
     clear_folder(ip)
     
-    # 加载动态链接库
-    try:
-        ctypes.CDLL(PLUGIN_LIBRARY)
-    except Exception as e:
-        logger.error(f"加载动态链接库失败: {e}")
-        sys.exit(1)
+    # # 加载动态链接库
+    # try:
+    #     ctypes.CDLL(PLUGIN_LIBRARY)
+    # except Exception as e:
+    #     logger.error(f"加载动态链接库失败: {e}")
+    #     sys.exit(1)
         
     # 手动输入训练时的类别
-    categories = ["shoe"]
+    categories = ["class1", "class2"]
     
     # 加载模型文件
     try:
-        yolov8_wrapper1 = YoLov8TRT(engine_file_path)
+        yolov8_wrapper1 = YOLOv8RKNN(rknn_path, categories)
     except Exception as e:
         logger.error(f"加载模型文件失败: {e}")
         sys.exit(1)

util/yolov8rknn.py

@@ -7,20 +7,14 @@ import time
 
 from rknn.api import RKNN
 
-
-
-
-
-    
-    
 class YOLOv8RKNN:
-    def __init__(self, model_path, input_size=(640, 640)):
+    def __init__(self, model_path, input_size=(640, 640), class_names = ['class1', 'class2']):
         self.model_path = model_path
         self.input_size = input_size
         self.rknn = RKNN()
         
         # 类别名称，根据你的2个类别修改
-        self.class_names = ['class1', 'class2']  # 请替换为你实际的类别名称
+        self.class_names = class_names  # 请替换为你实际的类别名称
         
         # 初始化模型
         self.load_model()
@@ -148,6 +142,21 @@ class YOLOv8RKNN:
         
         return boxes, scores, class_ids, inference_time
     
+    def infer(self, frame, conf_threshold=0.5, nms_threshold=0.4, testclasses=None):
+        # 预处理
+        input_image = self.preprocess(frame)
+        outputs = self.rknn.inference(inputs=[input_image])
+        boxes, scores, class_ids = self.postprocess(outputs, conf_threshold, nms_threshold)
+        if None:
+            return True, 0, []
+        result = list()
+        for i in range(class_ids):
+            if class_ids[i] in testclasses:
+                result.append(boxes[i])
+        
+        return True, len(result), result
+        
+    
     def draw_detections(self, image, boxes, scores, class_ids):
         """在图像上绘制检测结果"""
         for i in range(len(boxes)):