Stable + Confidence + Hysteresis

app.py

@@ -1,5 +1,5 @@
 # ============================================================
-# 🚦 Stage 3 — Wrong Direction Detection (Improved + Stable + Confidence)
+# 🚦 Stage 3 — Wrong Direction Detection (Stable + Confidence + Hysteresis)
 # ============================================================
 
 import os, cv2, json, tempfile, numpy as np, gradio as gr
@@ -32,8 +32,9 @@ class Track:
         self.history = []
         self.frames_seen = 0
         self.status = "OK"
-        self.status_history = []  # 🔸 maintain recent decision history
+        self.status_history = []
         self.confidence = 1.0
+        self.ema_sim = 1.0  # for exponential smoothing
 
     def update(self, bbox):
         self.kf.predict()
@@ -45,12 +46,11 @@ class Track:
         self.frames_seen += 1
         return [x, y]
 
-    def stable_status(self, new_status, new_conf, window=6, agree_ratio=0.7):
+    def stable_status(self, new_status, new_conf, window=10, agree_ratio=0.6):
         """Debounce flicker using recent window consensus."""
         self.status_history.append(new_status)
         if len(self.status_history) > window:
             self.status_history.pop(0)
-        # majority consensus
         if self.status_history.count(new_status) >= int(agree_ratio * len(self.status_history)):
             self.status = new_status
             self.confidence = new_conf
@@ -58,7 +58,7 @@ class Track:
 
 
 # ============================================================
-# ⚙️ Utilities
+# ⚙️ Utility Functions
 # ============================================================
 def compute_cosine_similarity(v1, v2):
     v1 = v1 / (np.linalg.norm(v1) + 1e-6)
@@ -91,9 +91,12 @@ def process_video(video_file, stage2_json, show_only_wrong=False):
     out = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
 
     tracks, next_id = {}, 0
-    SIM_THRESH = 0.5          # cosine similarity threshold
-    DELAY_FRAMES = 8          # wait N frames before flagging
-    MIN_FLOW_SPEED = 1.2      # ignore jitter
+    SIM_THRESH = 0.5          # base reference
+    DELAY_FRAMES = 8
+    MIN_FLOW_SPEED = 1.2
+    HYST_OK = 0.55
+    HYST_WRONG = 0.45
+    ALPHA = 0.6               # exponential smoothing weight
 
     while True:
         ret, frame = cap.read()
@@ -140,15 +143,23 @@ def process_video(video_file, stage2_json, show_only_wrong=False):
             if np.linalg.norm(motion) < MIN_FLOW_SPEED:
                 continue
 
-            # cosine similarity vs lane flows
             sims = [compute_cosine_similarity(motion, f) for f in lane_flows]
             best_sim = max(sims)
 
             if trk.frames_seen > DELAY_FRAMES:
-                new_status = "WRONG" if best_sim < SIM_THRESH else "OK"
-                trk.stable_status(new_status, new_conf=best_sim)
+                # Exponential moving average
+                trk.ema_sim = ALPHA * best_sim + (1 - ALPHA) * getattr(trk, "ema_sim", best_sim)
+
+                # Hysteresis classification
+                if trk.ema_sim >= HYST_OK:
+                    new_status = "OK"
+                elif trk.ema_sim <= HYST_WRONG:
+                    new_status = "WRONG"
+                else:
+                    new_status = trk.status  # hold previous label
+
+                trk.stable_status(new_status, new_conf=trk.ema_sim, window=10, agree_ratio=0.6)
 
-                # draw if toggle allows
                 if (not show_only_wrong) or (trk.status == "WRONG"):
                     color = (0, 0, 255) if trk.status == "WRONG" else (0, 255, 0)
                     label = f"ID:{tid} {trk.status} ({trk.confidence:.2f})"
@@ -165,13 +176,12 @@ def process_video(video_file, stage2_json, show_only_wrong=False):
 # 🎛️ Gradio Interface
 # ============================================================
 description = """
-### 🚦 Stage 3 — Wrong Direction Detection (Improved + Stable + Confidence)
-- ✅ Cosine similarity instead of raw angle  
-- ✅ Lane-wise flow support for curved roads  
-- ✅ Temporal smoothing & delayed classification  
-- ✅ Flicker-free labels using consensus window  
-- ✅ Confidence score display (cosine similarity)
-- ✅ Toggle to show only WRONG overlays  
+### 🚦 Stage 3 — Wrong Direction Detection (Stable + Confidence + Hysteresis)
+- ✅ Cosine similarity with exponential smoothing  
+- ✅ Hysteresis (OK≥0.55 / WRONG≤0.45) for stability  
+- ✅ 10-frame consensus voting (flicker-free)  
+- ✅ Confidence score beside each ID  
+- ✅ Optional “Show Only Wrong Labels” toggle
 """
 
 demo = gr.Interface(