add models

This commit is contained in:
Sebastian Unterschütz
2026-04-21 15:45:43 +02:00
parent f5d344cf75
commit e72ba2989e
14 changed files with 1589 additions and 79 deletions

View File

@@ -4,43 +4,69 @@ Dieses Projekt ist eine hochoptimierte KI-Steuerung für die DJI Tello Drohne (u
## 🚀 Hauptmerkmale ## 🚀 Hauptmerkmale
- **Hybrid AI Engine**: Nutzt **YuNet** (Face Detection), **MiDaS** (Depth Estimation) und **MobileNetV3** (ReID) alle via **ONNX Runtime** für maximale FPS. - **Hybrid AI Engine**: Nutzt **YuNet** (Face Detection), **MiDaS** (Depth Estimation) und **MobileNetV3** (ReID) alle via **ONNX Runtime** für maximale FPS.
- **Multithreading**: Kamera-Stream und KI-Verarbeitung laufen in getrennten Threads. Das Videobild bleibt flüssig, egal wie schwer die KI arbeitet. - **Sport-Modus**: Aggressives 3-Achsen-Tracking für schnelle Verfolgungsjagden.
- **Visual Fingerprinting (ReID)**: Die Drohne erstellt einen digitalen Fingerabdruck einer fixierten Person und findet sie automatisch wieder, wenn sie kurz aus dem Bild verschwindet. - **Visual Fingerprinting (ReID)**: Erkennt fixierte Personen wieder, auch wenn sie das Bild kurz verlassen.
- **Intelligente Verfolgung**: Aktive Suche in Verschwindungsrichtung und aggressiver Verfolgungsmodus. - **Multithreading**: Flüssiges Kamerabild durch Trennung von Video-Stream und KI-Logik.
- **Echtzeit-HUD**: Professionelles Display mit Telemetrie, AI-Tiefenkarte und Radar-Zonen zur Hindernisvermeidung. - **Simulator inklusive**: Eine integrierte Ursina-Umgebung zum gefahrlosen Testen.
- **Rate Limiting**: Kontrollierter Datenfluss (10Hz) zum Tello-SDK für maximale Verbindungsstabilität.
## 🕹 Steuerung ## 🛠 Installation
Klicke mit der **Maus** direkt in das Videobild, um eine Person zu fixieren (**Lock-ON**).
### Tastatur-Belegung: 1. **Python-Umgebung**: Empfohlen wird Python 3.10 bis 3.12.
2. **Abhängigkeiten installieren**:
```bash
pip install -r requirements.txt
```
3. **Modelle prüfen**: Stelle sicher, dass im Ordner `models/` folgende Dateien liegen:
- `face_detection_yunet.onnx`
- `midas_small.onnx`
- `reid_mobilenet.onnx`
## 🚀 Starten der Drohne
### A) Im Simulator (Standard)
1. Öffne die Datei `run.py`.
2. Stelle sicher, dass `use_real_tello=False` eingestellt ist.
3. Starte das Programm:
```bash
python run.py
```
4. Das Simulator-Fenster und das KI-Pilot-Fenster öffnen sich automatisch.
### B) Mit der echten DJI Tello
1. Schalte deine Tello ein und verbinde deinen PC mit dem WLAN der Drohne (z.B. `TELLO-XXXXXX`).
2. Öffne die Datei `run.py` und ändere die Zeile zu:
```python
app = FaceTrackingApp(use_real_tello=True)
```
3. Starte das Programm:
```bash
python run.py
```
## 🕹 Bedienung im Flug
Sobald das Video-Fenster erscheint:
1. **Klicke mit der Maus** in das Fenster, um den Fokus zu aktivieren.
2. Drücke **'T'** zum Starten (Takeoff).
3. **Tracking aktivieren**: Klicke mit der Maus auf ein erkanntes Gesicht/Person oder drücke **'K'**, um den nächsten erkannten Kopf zu fixieren (Lock-ON).
### Wichtige Tasten:
| Taste | Aktion | | Taste | Aktion |
| :--- | :--- | | :--- | :--- |
| **T** | **Takeoff** (Starten) - Einmal drücken und kurz warten | | **T** | **Takeoff** (Abheben) |
| **L** | **Land** (Landen) | | **L** | **Land** (Landen) |
| **M** | Wechsel zwischen **Manuell** und **KI-Modus** | | **2** | **Sport-Modus** (An/Aus) - Schnelleres Tracking auf allen Achsen |
| **K** | Lock-ON Trigger aktivieren / Fixierung löschen | | **M** | **Manueller Modus** (KI stoppt die Steuerung) |
| **Space**| **Not-Aus** (Stoppt alle Motoren sofort) | | **Space**| **Not-Aus** (Motoren sofort aus) |
| **W/S** | Vorwärts / Rückwärts (Manuell) | | **W/S/A/D**| Vorwärts, Rückwärts, Links, Rechts (Manuell) |
| **A/D** | Links / Rechts (Manuell) |
| **R/F** | Steigen / Sinken (Manuell) | | **R/F** | Steigen / Sinken (Manuell) |
| **E/Z** | Drehen Links / Rechts (Manuell) | | **E/Z** | Drehen Links / Rechts (Manuell) |
| **1** | Automatisches Drehen (Scan-Modus) an/aus | | **Enter**| Programm sicher beenden |
| **Enter**| Programm beenden |
## 🏗 Architektur ## 🏗 Architektur & Performance
- `drone_pilot/main.py`: Das Herzstück. Verwaltet Threads, UI-Events und die Hauptschleife. - **AI-Worker Thread**: Verarbeitet alle Modelle parallel zum Haupt-Thread.
- `drone_pilot/vision.py`: Die KI-Engine. Lädt ONNX-Modelle und verarbeitet Bilddaten. - **Rate Limiting**: RC-Befehle werden mit 10Hz gesendet, um das SDK-Protokoll nicht zu überlasten.
- `drone_pilot/flight.py`: Der Flug-Controller. Berechnet RC-Vektoren basierend auf KI-Ergebnissen. - **ONNX Acceleration**: Nutzt CPU-Optimierungen, die bis zu 5x schneller als Standard-PyTorch sind.
- `drone_pilot/ui.py`: HUD-Renderer für das OpenCV-Fenster.
- `drone_pilot/config.py`: Zentrale Konfiguration für Geschwindigkeiten und Schwellenwerte.
## 🛠 Installation & Start
1. Installiere die Abhängigkeiten: `pip install -r requirements.txt`
2. Stelle sicher, dass die Modelle im Ordner `models/` liegen.
3. Starte das Programm:
- Simulator: `python run.py` (Default: `use_real_tello=False`)
- Reale Drohne: Ändere in `run.py` zu `use_real_tello=True`
--- ---
*Hinweis: Beim Fliegen mit der realen Drohne immer auf einen Akkustand > 15% achten, da der Startbefehl sonst vom SDK abgelehnt wird.* *Viel Spaß beim Fliegen! Achte bei der echten Drohne immer auf genügend Platz und einen vollen Akku (>15%).*

View File

@@ -0,0 +1 @@
<?xml version='1.0' encoding='UTF-8'?><Error><Code>NoSuchKey</Code><Message>The specified key does not exist.</Message><Details>No such object: mediapipe-models/face_detector/blaze_face_long_range/float16/latest/blaze_face_long_range.tflite</Details></Error>

View File

@@ -1,6 +1,6 @@
# drone_pilot/config.py # drone_pilot/config.py
class Config: class Config:
WIN_NAME = "Tello AI Pilot v2.0 (High Speed)" WIN_NAME = "Tello AI Pilot v2.0 (Extreme Speed)"
WIDTH, HEIGHT = 1024, 720 WIDTH, HEIGHT = 1024, 720
TARGET_ALTITUDE = 1.5 TARGET_ALTITUDE = 1.5
@@ -8,21 +8,28 @@ class Config:
TARGET_PERSON_SIZE = 400 TARGET_PERSON_SIZE = 400
ALT_THRESHOLD = 0.12 ALT_THRESHOLD = 0.12
YAW_GAIN = 0.08 # Reduced for smoother rotation # Normal Mode Gains
FORWARD_GAIN = 1.5 # Kept high for fast pursuit YAW_GAIN = 0.12
FORWARD_GAIN = 1.5
ALT_GAIN = 40 ALT_GAIN = 40
# Sport Mode Gains (Much more aggressive)
SPORT_YAW_GAIN = 0.25
SPORT_FB_GAIN = 2.0
SPORT_LR_GAIN = 0.6
DEPTH_THRESHOLD = 0.90 DEPTH_THRESHOLD = 0.90
OBSTACLE_TOF_CM = 70 OBSTACLE_TOF_CM = 70
FACE_DEADZONE = 30 # Slightly larger deadzone for stability FACE_DEADZONE = 20
FACE_ROT_ONLY = 100 FACE_ROT_ONLY = 80
PERSON_CONF_THRESHOLD = 0.5 PERSON_CONF_THRESHOLD = 0.5
SMOOTHING_ALPHA = 0.35 # High directness, but slightly more damped than extreme SMOOTHING_ALPHA = 0.35
class Colors: class Colors:
GREEN = (0, 255, 0) GREEN = (0, 255, 0)
RED = (0, 0, 255) RED = (0, 0, 255)
BLUE = (255, 0, 0) BLUE = (255, 0, 0)
PURPLE = (255, 0, 255)
WHITE = (255, 255, 255) WHITE = (255, 255, 255)
BLACK = (0, 0, 0) BLACK = (0, 0, 0)
HUD_BG = (10, 10, 10) HUD_BG = (10, 10, 10)

View File

@@ -4,19 +4,24 @@ from typing import List, Tuple, Dict
from .config import Config from .config import Config
class FlightController: class FlightController:
"""
Verantwortlich für die Berechnung der Flugvektoren basierend auf KI-Ergebnissen.
Unterstützt Normal-Modus (sequenziell) und Sport-Modus (simultan/LR).
"""
def __init__(self): def __init__(self):
self.last_sent_rc = [0, 0, 0, 0] self.last_sent_rc = [0, 0, 0, 0]
self.smooth_face = None self.smooth_face = None
self.search_start = time.time() self.search_start = time.time()
self.status = "INITIALIZING" self.status = "INITIALIZING"
# Memory for lost targets # Speicher für verloren gegangene Ziele
self.last_target_side = 0 # -1 for left, 1 for right self.last_target_side = 0
self.lost_time = 0 self.lost_time = 0
def calculate(self, def calculate(self,
faces: List[Tuple], faces: List[Tuple],
is_manual: bool, is_manual: bool,
is_sport: bool, # NEU: Sport-Modus Flag
emergency_stop: bool, emergency_stop: bool,
is_locked: bool, is_locked: bool,
locked_person: Tuple, locked_person: Tuple,
@@ -29,7 +34,6 @@ class FlightController:
lr, fb, ud, yv = 0, 0, 0, 0 lr, fb, ud, yv = 0, 0, 0, 0
# Face smoothing for UI/Visuals
if len(faces) > 0: if len(faces) > 0:
target = max(faces, key=lambda f: f[2] * f[3]) target = max(faces, key=lambda f: f[2] * f[3])
if self.smooth_face is None: self.smooth_face = target if self.smooth_face is None: self.smooth_face = target
@@ -42,13 +46,13 @@ class FlightController:
self.status = "EMERGENCY STOP" self.status = "EMERGENCY STOP"
return (0, 0, 0, 0), self.status return (0, 0, 0, 0), self.status
# Obstacle Avoidance (always active if flying) # Hindernisvermeidung
center_blocked = zones["CENTER"] or tof < Config.OBSTACLE_TOF_CM center_blocked = zones["CENTER"] or tof < Config.OBSTACLE_TOF_CM
if center_blocked: if center_blocked:
self.status = "AVOIDING OBSTACLE" self.status = "AVOIDING OBSTACLE"
yv = 80 if zone_scores["LEFT"] < zone_scores["RIGHT"] else -80 yv = 80 if zone_scores["LEFT"] < zone_scores["RIGHT"] else -80
fb = -30 fb = -30
return self._smooth(lr, fb, ud, yv) return self._smooth(0, fb, 0, yv)
if is_manual: if is_manual:
self.status = "MANUAL CONTROL" self.status = "MANUAL CONTROL"
@@ -59,54 +63,57 @@ class FlightController:
# AI LOGIC # AI LOGIC
if is_locked: if is_locked:
if locked_person is not None: if locked_person is not None:
# Target is visible -> Normal Pursuit
self.search_start = time.time() self.search_start = time.time()
self.lost_time = 0 self.lost_time = 0
(x, y, w, h) = locked_person (x, y, w, h) = locked_person
center_x = x + w // 2 center_x = x + w // 2
err_x = center_x - (Config.WIDTH // 2) err_x = center_x - (Config.WIDTH // 2)
# Remember which side it was on
self.last_target_side = 1 if err_x > 0 else -1 self.last_target_side = 1 if err_x > 0 else -1
# Rotation (Yaw) - SMOOTHER if is_sport:
# SPORT MODUS: Alles gleichzeitig + LR-Strafing
yv = int(np.clip(Config.SPORT_YAW_GAIN * err_x, -100, 100))
fb = int(np.clip(Config.SPORT_FB_GAIN * (Config.TARGET_PERSON_SIZE - w), -100, 100))
lr = int(np.clip(Config.SPORT_LR_GAIN * err_x, -60, 60))
self.status = "SPORT PURSUIT: FULL AXIS"
else:
# NORMAL MODUS: Sequenziell (Drehen ODER Fliegen)
if abs(err_x) > Config.FACE_DEADZONE: if abs(err_x) > Config.FACE_DEADZONE:
yv = int(np.clip(Config.YAW_GAIN * err_x, -50, 50)) yv = int(np.clip(Config.YAW_GAIN * err_x, -50, 50))
fb = 0
# Forward/Backward pursuit - EXTREME SPEED self.status = "PURSUIT: AIMING"
alignment_factor = max(0.4, 1.0 - (abs(err_x) / Config.FACE_ROT_ONLY))
target_fb = int(np.clip(Config.FORWARD_GAIN * (Config.TARGET_PERSON_SIZE - w), -90, 90))
fb = int(target_fb * alignment_factor)
self.status = "PURSUIT: EXTREME"
else: else:
# Target is LOST -> Rapid Search logic yv = 0
fb = int(np.clip(Config.FORWARD_GAIN * (Config.TARGET_PERSON_SIZE - w), -80, 80))
self.status = "PURSUIT: APPROACHING"
else:
# Target verloren
if self.lost_time == 0: self.lost_time = time.time() if self.lost_time == 0: self.lost_time = time.time()
elapsed = time.time() - self.lost_time elapsed = time.time() - self.lost_time
if elapsed < 10.0: # Search longer and faster search_speed = 60 if is_sport else 40
yv = 40 * self.last_target_side if elapsed < 10.0:
yv = search_speed * self.last_target_side
self.status = f"LOST TARGET: SCANNING {'RIGHT' if self.last_target_side > 0 else 'LEFT'}" self.status = f"LOST TARGET: SCANNING {'RIGHT' if self.last_target_side > 0 else 'LEFT'}"
else: else:
self.status = "TARGET LOST: AGGRESSIVE PATROL" self.status = "TARGET LOST: PATROL"
yv = 30 yv = 30
elif self.smooth_face is not None: elif self.smooth_face is not None:
# Face found but not locked
(x, y, w, h) = self.smooth_face (x, y, w, h) = self.smooth_face
err_x = (x + w // 2) - (Config.WIDTH // 2) err_x = (x + w // 2) - (Config.WIDTH // 2)
if abs(err_x) > Config.FACE_DEADZONE:
yv = int(np.clip(Config.YAW_GAIN * err_x, -40, 40)) yv = int(np.clip(Config.YAW_GAIN * err_x, -40, 40))
self.status = "AWAITING LOCK" self.status = "AWAITING LOCK"
else: else:
# Patrol mode - faster # Patrouille
elapsed = (time.time() - self.search_start) % 6.0 elapsed = (time.time() - self.search_start) % 8.0
if elapsed < 2.0: if elapsed < 3.0:
self.status = "PATROL: DASH" self.status = "PATROL: ADVANCE"
fb = 40 fb = 35
else: else:
self.status = "PATROL: SCAN" self.status = "PATROL: SCAN"
yv = 30 yv = 35
return self._smooth(lr, fb, ud, yv) return self._smooth(lr, fb, ud, yv)
@@ -117,10 +124,10 @@ class FlightController:
sud = int(self.last_sent_rc[2] * (1-alpha) + ud * alpha) sud = int(self.last_sent_rc[2] * (1-alpha) + ud * alpha)
syv = int(self.last_sent_rc[3] * (1-alpha) + yv * alpha) syv = int(self.last_sent_rc[3] * (1-alpha) + yv * alpha)
if abs(slr) < 2: slr = 0 if abs(slr) < 3: slr = 0
if abs(sfb) < 2: sfb = 0 if abs(sfb) < 3: sfb = 0
if abs(sud) < 2: sud = 0 if abs(sud) < 3: sud = 0
if abs(syv) < 2: syv = 0 if abs(syv) < 3: syv = 0
self.last_sent_rc = [slr, sfb, sud, syv] self.last_sent_rc = [slr, sfb, sud, syv]
return (slr, sfb, sud, syv), self.status return (slr, sfb, sud, syv), self.status

View File

@@ -23,6 +23,7 @@ class FaceTrackingApp:
# State Management # State Management
self.is_running = True self.is_running = True
self.is_manual = True self.is_manual = True
self.is_sport = False # NEW: Sport Mode flag
self.is_locked = False self.is_locked = False
self.is_taking_off = False self.is_taking_off = False
self.is_flying = False self.is_flying = False
@@ -180,9 +181,10 @@ class FaceTrackingApp:
self.locked_person_features = feat self.locked_person_features = feat
self.is_locked = True self.is_locked = True
self.lock_trigger = False self.lock_trigger = False
self.is_manual = False
break break
# UI # UI Graphics
if self.is_locked and self.locked_person: if self.is_locked and self.locked_person:
(x,y,w,h) = self.locked_person (x,y,w,h) = self.locked_person
cv2.rectangle(frame, (x,y), (x+w,y+h), Colors.BLUE, 3) cv2.rectangle(frame, (x,y), (x+w,y+h), Colors.BLUE, 3)
@@ -204,6 +206,7 @@ class FaceTrackingApp:
rc, status = self.flight_controller.calculate( rc, status = self.flight_controller.calculate(
faces=faces, faces=faces,
is_manual=self.is_manual, is_manual=self.is_manual,
is_sport=self.is_sport, # NEW
emergency_stop=self.emergency_stop, emergency_stop=self.emergency_stop,
is_locked=self.is_locked, is_locked=self.is_locked,
locked_person=self.locked_person, locked_person=self.locked_person,
@@ -215,7 +218,8 @@ class FaceTrackingApp:
manual_rc=active_manual_rc manual_rc=active_manual_rc
) )
# Throttle and optimized sending if self.is_sport: status = "SPORT MODE: ACTIVE"
now = time.time() now = time.time()
if now - self.last_rc_time >= 0.1: if now - self.last_rc_time >= 0.1:
changed = any(abs(rc[i] - self._prev_rc[i]) > 1 for i in range(4)) changed = any(abs(rc[i] - self._prev_rc[i]) > 1 for i in range(4))
@@ -250,7 +254,6 @@ class FaceTrackingApp:
def _handle_takeoff(self): def _handle_takeoff(self):
if self.is_taking_off or self.is_flying: return if self.is_taking_off or self.is_flying: return
self.is_taking_off = True self.is_taking_off = True
self.takeoff_error = False self.takeoff_error = False
def _task(): def _task():
@@ -265,13 +268,13 @@ class FaceTrackingApp:
self.takeoff_error = True self.takeoff_error = True
finally: finally:
self.is_taking_off = False self.is_taking_off = False
threading.Thread(target=_task, daemon=True).start() threading.Thread(target=_task, daemon=True).start()
def _handle_input(self, key: int): def _handle_input(self, key: int):
if key == 13: self.is_running = False if key == 13: self.is_running = False
elif key == 32: self.emergency_stop = not self.emergency_stop elif key == 32: self.emergency_stop = not self.emergency_stop
elif key == ord('m'): self.is_manual = not self.is_manual elif key == ord('m'): self.is_manual = not self.is_manual
elif key == ord('2'): self.is_sport = not self.is_sport # Toggle Sport Mode
elif key == ord('k'): elif key == ord('k'):
self.lock_trigger = not self.lock_trigger self.lock_trigger = not self.lock_trigger
self.is_locked = False self.is_locked = False
@@ -287,11 +290,9 @@ class FaceTrackingApp:
except: pass except: pass
elif key == ord('1'): self.is_rotating = not self.is_rotating elif key == ord('1'): self.is_rotating = not self.is_rotating
# Reset manual speed
self.m_lr, self.m_fb, self.m_ud, self.m_yv = 0, 0, 0, 0 self.m_lr, self.m_fb, self.m_ud, self.m_yv = 0, 0, 0, 0
if self.is_manual and not self.emergency_stop: if self.is_manual and not self.emergency_stop:
s = 100 # Maximum manual speed s = 100
if key == ord('w'): self.m_fb = s if key == ord('w'): self.m_fb = s
elif key == ord('s'): self.m_fb = -s elif key == ord('s'): self.m_fb = -s
elif key == ord('a'): self.m_lr = -s elif key == ord('a'): self.m_lr = -s

File diff suppressed because one or more lines are too long

BIN
models/midas_small.onnx Normal file

Binary file not shown.

Binary file not shown.

BIN
models/mobilenet_v3.onnx Normal file

Binary file not shown.

Binary file not shown.

BIN
models/reid_mobilenet.onnx Normal file

Binary file not shown.

Binary file not shown.

0
models/yolov8n.onnx Normal file
View File

View File

@@ -1,4 +1,10 @@
ursina==7.0.0 ursina==7.0.0
PyOpenGL==3.1.9 PyOpenGL==3.1.9
numpy==2.2.3 numpy>=2.0.0
opencv-python==4.11.0.86 opencv-python>=4.10.0.0
opencv-contrib-python>=4.10.0.0
djitellopy>=2.5.0
onnxruntime>=1.19.0
torch>=2.0.0
torchvision>=0.15.0
Pillow>=10.0.0