feat: configurations

geolocation.py

@@ -32,7 +32,6 @@ class Geolocation:
         # Вычисляем угол поворота
         rotation = -np.arctan2(mat[1, 0], mat[0, 0])
         scale = np.hypot(mat[0, 0], mat[0, 1])
-        print("Scale: ", scale)
         
         # Координаты уже отцентрированы, поэтому используем их напрямую
         dx_meters = tx

main.py

@@ -34,7 +34,7 @@ def main():
     # points = get_trajectory_points('map.jpg')
 
     # Trajectory #1
-    points = [[np.float64(0.5384504359393909), np.float64(0.4084520767967683)], [np.float64(0.4451750568707629), np.float64(0.38213330305374654)], [np.float64(0.49266070439660997), np.float64(0.2789637099811013)], [np.float64(0.36377108968359656), np.float64(0.3263375027185404)], [np.float64(0.3535955937852008), np.float64(0.4337180995900692)]]
+    # points = [[np.float64(0.5384504359393909), np.float64(0.4084520767967683)], [np.float64(0.4451750568707629), np.float64(0.38213330305374654)], [np.float64(0.49266070439660997), np.float64(0.2789637099811013)], [np.float64(0.36377108968359656), np.float64(0.3263375027185404)], [np.float64(0.3535955937852008), np.float64(0.4337180995900692)]]
 
     # Trajectory #2
     # points = [[np.float64(0.29197731306713737), np.float64(0.3452870198135161)], [np.float64(0.33494051797147517), np.float64(0.2010601397017569)], [np.float64(0.39768940934491587), np.float64(0.25369768718780034)], [np.float64(0.4027771572941138), np.float64(0.4158213334448144)], [np.float64(0.2914120077394487), np.float64(0.5547844588079692)]]
@@ -93,6 +93,7 @@ def main():
         kp, des = chunk.compute_keypoints()
 
         plt.subplot(1, len(points), i+1)
+        plt.imshow(chunk.image)
         kp_coords = np.array([j.pt for j in kp])
         if len(kp_coords) > 0:
             plt.scatter(kp_coords[:, 0], kp_coords[:, 1], c='red', s=20, alpha=0.7, marker='o')
@@ -187,3 +188,5 @@ def main():
 
 if __name__ == "__main__":
     main()
+
+#TODO

todo.md

@@ -18,3 +18,7 @@
 | [+] Уменьшение погрешность за счёт удаления выбросов
 | [+] Исследовать причину погрешности при развороте
 [+] Устранение большой погрешности при повороте
+
+[ ] Переделать ключевые точки -> Optical Flow
+[ ] Добавить перспективу
+[ ] Эталоны на Google Maps, полёт тот же

vision_chunk.py

@@ -6,12 +6,13 @@ from pathlib import Path
 from typing import Literal, Optional, Tuple
 from PIL import Image
 
-FeatureMethod = Literal["orb", "sift", "surf", "akaze", "brisk"]
+FeatureMethod = Literal["orb", "sift", "akaze", "brisk"]
+DEFAULT_METHOD = "akaze"
 
 @dataclass
 class VisionChunk:
     image: Image.Image
-    feature_method: FeatureMethod = "orb"
+    feature_method: FeatureMethod = DEFAULT_METHOD
 
     keypoints: Optional[list] = field(default=None, init=False)
     descriptors: Optional[np.ndarray] = field(default=None, init=False)
@@ -26,7 +27,7 @@ class VisionChunk:
             self._detector = cv2.ORB_create(
                 nfeatures=1000,
                 scaleFactor=1.2,
-                nlevels=8,
+                nlevels=32,
                 edgeThreshold=31,
                 firstLevel=0,
                 WTA_K=2,
@@ -35,18 +36,11 @@ class VisionChunk:
             )
         elif self.feature_method == "sift":
             self._detector = cv2.SIFT_create(
-                nfeatures=1000,
-                nOctaveLayers=3,
-                contrastThreshold=0.04,
-                edgeThreshold=10,
-                sigma=1.6
-            )
-        elif self.feature_method == "surf":
-            self._detector = cv2.xfeatures2d.SURF_create(
-                hessianThreshold=400.0,
-                nOctaveLayers=3,
-                # nOctaveLayers=4,
-                upright=False
+                nfeatures=1500,
+                nOctaveLayers=2,
+                contrastThreshold=0.01,
+                edgeThreshold=15,
+                sigma=3.3
             )
         elif self.feature_method == "akaze":
             self._detector = cv2.AKAZE_create(
@@ -59,8 +53,8 @@ class VisionChunk:
             )
         elif self.feature_method == "brisk":
             self._detector = cv2.BRISK_create(
-                thresh=30,
-                octaves=3,
+                thresh=70,
+                octaves=7,
                 patternScale=1.0
             )
         else:
@@ -69,15 +63,10 @@ class VisionChunk:
 
     def _get_matcher(self) -> cv2.DescriptorMatcher:
         if self._matcher is None:
-            self._matcher = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=False)
-            return self._matcher
-            if self.feature_method == 'orb':
+            if self.feature_method == 'sift':
+                self._matcher = cv2.BFMatcher(cv2.NORM_L2, crossCheck=False)
+            else:
                 self._matcher = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=False)
-            elif self.feature_method == 'sift':
-                FLANN_INDEX_LSH = 6
-                index_params = dict(algorithm=FLANN_INDEX_LSH, table_number=6, key_size=12, multi_probe_level=1)
-                search_params = dict(checks=50)
-                self._matcher = cv2.FlannBasedMatcher(index_params, search_params)
         return self._matcher
 
     def _preprocess(self, img_np: np.ndarray) -> np.ndarray:
@@ -90,7 +79,7 @@ class VisionChunk:
         clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
         return clahe.apply(gray)
 
-    def compute_keypoints(self, force: bool = False) -> Tuple[list, Optional[np.ndarray]]:
+    def compute_keypoints(self, force: bool = False) -> Tuple[list[cv2.KeyPoint], Optional[np.ndarray]]:
         if self.keypoints is not None and self.descriptors is not None and not force:
             return self.keypoints, self.descriptors
 
@@ -103,11 +92,21 @@ class VisionChunk:
         
         # CLAHE предобработка
         preprocessed = self._preprocess(img_np)
-        kps, desc = detector.detectAndCompute(preprocessed, None)
+        keypoints, descriptors = detector.detectAndCompute(preprocessed, None)
 
-        self.keypoints = kps
-        self.descriptors = desc
-        return kps, desc
+        # Получаем массив response для всех точек
+        responses = np.array([kp.response for kp in keypoints])
+
+        # Находим индексы топ-100
+        top_indices = np.argsort(responses)[-2500:][::-1]
+
+        # Отбираем keypoints и descriptors
+        best_keypoints = [keypoints[i] for i in top_indices]
+        best_descriptors = descriptors[top_indices]
+
+        self.keypoints = best_keypoints
+        self.descriptors = best_descriptors
+        return self.keypoints, self.descriptors
 
     def detect_and_match_keypoints(
         self, 
@@ -193,7 +192,7 @@ class VisionChunk:
         return np.array(self.image)
 
     @classmethod
-    def load_image(cls, path: Path | str, feature_method: FeatureMethod = "orb") -> "VisionChunk":
+    def load_image(cls, path: Path | str, feature_method: FeatureMethod = DEFAULT_METHOD) -> "VisionChunk":
         path = Path(path)
         image = Image.open(path)
         return cls(image=image, feature_method=feature_method)