পাঠ ৩৩ · ৩৫-এর মধ্যে · মডিউল ৪
Home / AI Courses / Computer Vision / Project: realtime detection

প্রজেক্ট: real-time object detection

Project: real-time object detection — webcam YOLO end-to-end
১৫ মিনিট পড়া মাঝারি · Intermediate Project

এই project-এ যা করবেন

  • Bangladesh traffic dataset assemble (rickshaw, CNG, bus)
  • YOLOv8 finetune
  • Webcam realtime inference
  • Mobile deploy (ONNX export)

১ · Project scope

Goal: Bangladesh traffic-specific object detector — 5 class:

  • Rickshaw
  • CNG (auto)
  • Bus
  • Motorcycle
  • Pedestrian
Project pipeline

Real CV project — academic example beyond। Data collect, label, train, optimize, deploy, monitor — full lifecycle। Bangladesh-specific = local value।

২ · Step 1: Data collection

  • Source: dashcam video, traffic CCTV, smartphone, public footage।
  • Diversity: day/night, weather, location।
  • Quantity: 1000-5000 image initially।
  • Privacy: face/plate blur if public release।

Frame extraction:

import cv2
cap = cv2.VideoCapture('traffic_video.mp4')
i = 0
while True:
    ret, frame = cap.read()
    if not ret: break
    if i % 30 == 0:  # every 30 frame (1 sec at 30fps)
        cv2.imwrite(f'frames/{i:06d}.jpg', frame)
    i += 1
cap.release()

৩ · Step 2: Annotation

  • Tool: Roboflow, CVAT, LabelImg।
  • Format: YOLO (txt: class x_center y_center w h normalized)।
  • SAM-assist: click → mask → bbox auto extract।
  • Speed: 5-10 sec per image with SAM।

Dataset structure:

traffic_dataset/
├── images/
│   ├── train/   (80%)
│   ├── val/     (10%)
│   └── test/    (10%)
└── labels/
    ├── train/
    ├── val/
    └── test/

# data.yaml
path: ./traffic_dataset
train: images/train
val: images/val
names:
  0: rickshaw
  1: cng
  2: bus
  3: motorcycle
  4: pedestrian

৪ · Step 3: Train YOLOv8

Python · Ultralytics 
from ultralytics import YOLO

# Start from pretrained
model = YOLO('yolov8n.pt')  # nano — fastest

# Train
results = model.train(
    data='traffic_dataset/data.yaml',
    epochs=100,
    imgsz=640,
    batch=16,
    device=0,                       # GPU 0
    project='runs/train',
    name='bd_traffic',
    patience=20,                    # early stop
    cos_lr=True,                    # cosine LR schedule
    augment=True,                   # default augmentation
    mixup=0.1,                      # MixUp
    mosaic=1.0,                     # Mosaic 4-image
)

# Validate
metrics = model.val()
print(f"mAP@0.5: {metrics.box.map50:.3f}")
print(f"mAP@0.5:0.95: {metrics.box.map:.3f}")

    


    

      Training time estimate:
    

    
    
      
  • RTX 3060 (12 GB): 1000 image, 100 epoch — 2-3 hour।
    • 
      
  • Cloud T4 (Colab free): similar।
    • 
      
  • RTX 4090: 30 minute।
    • 
    


    
৫ · Step 4: Realtime webcam inference


    

      

        
        Python · OpenCV + YOLO
      

      
import cv2
from ultralytics import YOLO

# Load trained model
model = YOLO('runs/train/bd_traffic/weights/best.pt')

# Webcam (0 = default camera)
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 1280)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)

while True:
    ret, frame = cap.read()
    if not ret: break

    # Inference
    results = model(frame, verbose=False, conf=0.5)

    # Visualize
    annotated = results[0].plot()

    # FPS
    fps = 1 / (results[0].speed['inference'] / 1000)
    cv2.putText(annotated, f'FPS: {fps:.1f}', (10, 30),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)

    cv2.imshow('BD Traffic Detection', annotated)
    if cv2.waitKey(1) & 0xFF == ord('q'): break

cap.release()
cv2.destroyAllWindows()

      

        
        
      

      

    


    
৬ · Step 5: Optimize for deployment


    
(ক) ONNX export

    
model.export(format='onnx', imgsz=640, optimize=True)
# Creates yolov8n.onnx — framework-independent


    
(খ) Quantization (FP16)

    
model.export(format='onnx', half=True)
# 50% smaller file, 2x faster on supported hardware


    
(গ) TensorRT (NVIDIA)

    
model.export(format='engine', half=True)
# Maximum speed on NVIDIA GPU


    
(ঘ) Mobile (TFLite)

    
model.export(format='tflite', int8=True)
# Android/iOS deployment


    
৭ · Step 6: Bangladesh-specific tuning

    
    
      
  • Bengali label: "rickshaw" → "রিকশা" — UI display।
    • 
      
  • Local context: Pohela Boishakh decoration vehicle, Eid eve traffic।
    • 
      
  • Lighting: Dhaka monsoon, summer heat haze augmentation।
    • 
      
  • Camera: diverse smartphone — augment quality।
    • 
    


    
৮ · Step 7: Counting + tracking

    

      Detection alone insufficient for "how many cars passed?"। Need tracking।
    


    
# Built-in tracking with BotSORT/ByteTrack
results = model.track(
    source='video.mp4',
    tracker='botsort.yaml',
    persist=True
)
# Each box has unique ID across frames

# Count by class crossing line
line_y = 400
counts = {}
for r in results:
    for box, cls, id in zip(r.boxes.xyxy, r.boxes.cls, r.boxes.id):
        cy = (box[1] + box[3]) / 2
        if cy > line_y and id not in counts:
            counts[id.item()] = cls.item()


    

      Real CV project = "data plumbing > model architecture"। 80% effort: data collect, label, augment, evaluate। 20%: model train। Production = 99% engineering, 1% magic।
    


    
    

      
        
          
            
          
        
        Real-time detection — full project pipeline

        
        📹 Collect
        video frames

        
        🏷 Annotate
        SAM-assist

        
        🎯 Train
        YOLOv8

        
        📊 Validate
        mAP, precision

        
        ⚡ Optimize
        ONNX, INT8

        
        🚀 Deploy
        edge/cloud

        
        
        
        
        

        Iterative — each step feedback to previous
        Production model continuously evolve
        Monitor → identify failure case → augment dataset → retrain
      
      
Real CV project pipeline — collect, annotate, train, validate, optimize, deploy। Iterative।

    


    
৯ · Common pitfall

    
    
      
  • Class imbalance: rickshaw 1000 image, bus 100 — bus accuracy poor। Resample বা weight loss।
    • 
      
  • Domain shift: day train, night fail। Augment heavy।
    • 
      
  • Annotation noise: bbox loose — model learn loose। Strict QC।
    • 
      
  • Overfitting: small dataset — heavy regularization, transfer learning critical।
    • 
      
  • Realtime ambition: too-large model — drop FPS। Right-size choose।
    • 
    


    
১০ · Bangladesh-specific deployment

    
    
      
  • Smart traffic: Dhaka traffic management।
    • 
      
  • Vehicle counting: bridge, intersection।
    • 
      
  • Parking automation: shopping mall।
    • 
      
  • Safety: helmet detection (motorcyclist)।
    • 
      
  • Logistics: warehouse package detect।
    • 
    


    
১১ · Hardware options

    

      
        
HardwareFPS (YOLOv8n)CostUse

      
      
        
RTX 4090500+$1500Server

        
RTX 3060200$300Workstation

        
Jetson Nano15$100Edge

        
Raspberry Pi 55$80IoT

        
iPhone 15 Pro30flagshipMobile

      
    


    
১২ · Monitor in production

    
    
      
  • Confidence drift: avg confidence drop → model degrade।
    • 
      
  • Class distribution shift: new vehicle type emerging।
    • 
      
  • Failure case capture: low-confidence box log — manual review।
    • 
      
  • Retraining cycle: monthly/quarterly with new data।
    • 
    


    

      Real project — model accuracy 50% effort। Data, deployment, monitoring — others 50%। Beginner-এর mistake: too much focus on architecture। Lesson: pipeline > model।
    


  


  

    
    
  


  

    
 ভাবনার প্রশ্ন


    

      
        প্র ০১
        Project-এ "data quality" "data quantity"-এর চেয়ে বেশি matter। কেন? Examples?
      
      

        
Andrew Ng popularized "data-centric AI" — quality over quantity।

        
Quality issues:

        
    
          
  • Loose bbox — model learn loose।
    • 
          
  • Inconsistent label — class boundary blurry।
    • 
          
  • Mislabeled — direct error।
    • 
          
  • Missing annotation — false negative learn।
    • 
        

        
Quantity vs quality experiment:

        
    
          
  • 10K noisy image: 65% mAP।
    • 
          
  • 2K clean image: 75% mAP।
    • 
          
  • Better quality > 5x quantity।
    • 
        

        
Quality strategies:

        
    
          
  • Annotation guideline document।
    • 
          
  • Inter-annotator agreement check।
    • 
          
  • Random spot QC।
    • 
          
  • Active learning — focus on uncertain samples।
    • 
        

        
Bangladesh implication:

        
    
          
  • Local annotator training।
    • 
          
  • Bilingual guideline (Bangla + English)।
    • 
          
  • Cultural context understand।
    • 
        

        
মূল উপলব্ধি: "Garbage in, garbage out" — eternal। Investment in clean data — highest ROI।

      

    


    

      
        প্র ০২
        YOLOv8n vs YOLOv8x — accuracy vs speed। Bangladesh CCTV traffic-এ কোনটা?
      
      

        
Realtime production decision।

        
YOLOv8n (nano):

        
    
          
  • 3.2M params, 37% mAP।
    • 
          
  • 280 FPS T4।
    • 
          
  • Mobile deployable।
    • 
        

        
YOLOv8x (extra large):

        
    
          
  • 68M params, 53.9% mAP।
    • 
          
  • 40 FPS T4।
    • 
          
  • Server-only।
    • 
        

        
CCTV consideration:

        
    
          
  • 1080p 30 FPS stream।
    • 
          
  • Multiple camera (10-100)।
    • 
          
  • Server scaling matter।
    • 
          
  • Budget — RTX 3060 server bracket।
    • 
        

        
Decision matrix:

        
    
          
  • Single camera + accuracy priority: YOLOv8m/l।
    • 
          
  • 10+ camera + realtime: YOLOv8n/s।
    • 
          
  • Mobile patrol: YOLOv8n।
    • 
          
  • Forensic offline: YOLOv8x।
    • 
        

        
Cascade approach:

        
    
          
  • YOLOv8n always — fast first pass।
    • 
          
  • Low confidence → YOLOv8x verify।
    • 
          
  • Hybrid speed + accuracy।
    • 
        

        
Bangladesh real:

        
    
          
  • Dhaka City Corp pilot — YOLOv8s on 4 GPU server।
    • 
          
  • 4 camera per GPU।
    • 
          
  • Cost: $5K hardware + $1K/year maintenance।
    • 
        

        
মূল উপলব্ধি: "Best model" task-dependent। Speed/accuracy trade-off explicit reason।

      

    


    

      
        প্র ০৩
        Annotation cost reduce — SAM, active learning, synthetic data। Bangladesh-এ কীভাবে combine?
      
      

        
Bangladesh annotation industry opportunity।

        
SAM-assisted:

        
    
          
  • Click → mask → bbox extract।
    • 
          
  • Annotation 5-10x faster।
    • 
          
  • Quality often higher than manual।
    • 
        

        
Active learning:

        
    
          
  1. Initial 200 manual annotation।
    2. 
          
  2. Train baseline।
    3. 
          
  3. Predict on 5000 unlabeled।
    4. 
          
  4. Identify low-confidence — annotator priority।
    5. 
          
  5. Retrain — accuracy bump।
    6. 
          
  6. Iterate।
    7. 
        

        
Synthetic data:

        
    
          
  • Dhaka 3D city model render — annotated automatically।
    • 
          
  • Stable Diffusion + bbox — generative augment।
    • 
          
  • Domain gap — careful sim2real।
    • 
        

        
Combined strategy:

        
    
          
  • Initial: synthetic 5K।
    • 
          
  • Train baseline।
    • 
          
  • SAM-assist label real 1K।
    • 
          
  • Active learning identify hard real cases।
    • 
          
  • Annotator focus on hard, quality high।
    • 
          
  • Final dataset: 10K mixed।
    • 
        

        
Cost:

        
    
          
  • Pure manual: $1000।
    • 
          
  • SAM-assisted: $200।
    • 
          
  • Active learning: $100।
    • 
          
  • Combined: $50।
    • 
        

        
Bangladesh business:

        
    
          
  • Annotation export competitive।
    • 
          
  • Local rate + AI-tool — global service।
    • 
          
  • Skill: SAM operation, Python QC।
    • 
        

        
মূল উপলব্ধি: Modern annotation = AI + human collaboration। Bangladesh — labor cost advantage + AI tool literacy = lucrative niche।

      

    


    

      
        প্র ০৪
        Production deployment-এ MLOps — what minimum infrastructure for traffic detection in Dhaka?
      
      

        
MLOps Bangladesh context — affordable starter।

        
Minimum components:

        
    
          
  • Compute: 1 GPU server (RTX 3060)।
    • 
          
  • Camera: 5-10 IP cameras (Hikvision)।
    • 
          
  • Network: dedicated bandwidth।
    • 
          
  • Storage: 1TB for video archive।
    • 
        

        
Software stack:

        
    
          
  • Inference server: Triton or FastAPI।
    • 
          
  • Stream processing: OpenCV + Kafka (event)।
    • 
          
  • Database: PostgreSQL (detection log)।
    • 
          
  • Monitor: Prometheus + Grafana।
    • 
          
  • Dashboard: Streamlit বা React।
    • 
        

        
Pipeline:

        
    
          
  1. Camera → RTSP stream।
    2. 
          
  2. Worker reads frame।
    3. 
          
  3. YOLO inference।
    4. 
          
  4. Post-process: count, alert।
    5. 
          
  5. DB log।
    6. 
          
  6. Dashboard visualize।
    7. 
        

        
Deployment:

        
    
          
  • Docker Compose initial।
    • 
          
  • Kubernetes scale।
    • 
          
  • Cloud (GCP, AWS) BD startup-এ DigitalOcean।
    • 
        

        
Monitoring:

        
    
          
  • Inference latency p50/p99।
    • 
          
  • FPS per camera।
    • 
          
  • Detection confidence distribution drift।
    • 
          
  • Alert: latency > threshold, FPS drop।
    • 
        

        
Retraining:

        
    
          
  • Failure case capture (low-conf, error)।
    • 
          
  • Weekly review।
    • 
          
  • Monthly retrain।
    • 
          
  • A/B test new model।
    • 
        

        
Cost estimate (Bangladesh, 10 camera):

        
    
          
  • Hardware setup: $3000-5000।
    • 
          
  • Camera: $1500-3000।
    • 
          
  • Annual: $500 (electricity, maintenance)।
    • 
          
  • Engineer: 1 part-time।
    • 
        

        
Local consideration:

        
    
          
  • Power outage — UPS, generator।
    • 
          
  • Heat — cooling।
    • 
          
  • Theft — physical security।
    • 
          
  • Internet — fallback।
    • 
        

        
মূল উপলব্ধি: Production-grade CV — under $10K achievable Bangladesh। Government, private — ROI quick (vs manual operator cost)।

      

    

  


  

    
 অনুশীলন

    
    
      
  1. 
        Mini dataset: 50 image of household items, label, train YOLOv8n।
        
    
          
          
    
            
    Code section ৪ template। 30 minute Colab GPU।
    
          
    
        
    
      
    2. 

      
  2. 
        Webcam test: Pretrained YOLOv8 দিয়ে webcam realtime detect।
        
    
          
          
    
            
    Code section ৫। 80 COCO class — कुकुर, person, etc detect।
    
          
    
        
    
      
    3. 

      
  3. 
        ভাবুন: Bangladesh garment factory defect — YOLOv8 কীভাবে adapt?
        
    
          
          
    
            
    (1) Defect type 5-10। (2) Controlled lighting fixture। (3) Annotation standardize। (4) YOLOv8s (accuracy)। (5) Quality control workflow integrate।
    
          
    
        
    
      
    4. 
    

  


  

    
 আরও পড়ুন · ABCL TECH-এ আপনার পরবর্তী পদক্ষেপ

    
    

      কোড রানার কাজ না করলে? Google Colab use করুন।
    

  







  
   এই পাঠে ভুল পেয়েছেন? জানান





  
    
 পূর্ববর্তী পাঠ

    
পাঠ ৩২ · Pose estimation