Student Behavior Detection — YOLOv8 with Genetic Algorithm Optimization

[Audio] Assalamu Alaikum! My name is Shaheer, holding roll no 54 and my partner is Syed Ali Najaf Zaidi, holding roll no 63. Aaj hum aapke samne hamara research project present karne ja rahy hain, jiska title hai: Student Engagement Behavior Recognition using YOLOv8 optimized with a Genetic Algorithm. hamara AI system classroom ke 13 alag-alag student behaviors ko pehchanta hai. Humne model ko 50 epochs tak train kiya, jisse isne 98.6% ki kamaal ki accuracy haasil ki. Aur model ko mazeed automatic tune karne ke liye humne Genetic Algorithm ko 50 martaba chalaya, jahan hamen 0.565 ka sabsay best fitness score mila.

[Audio] Problem: Classroom ko manually check karna mushkil aur inefficient hai, aur purane AI models overhead view se student behaviors ko sahi tarah detect nahi kar paate. Solution: Humne lightweight aur fast YOLOv8n model use kiya, jise custom dataset par 50 epochs tak train kiya aur Genetic Algorithm (GA) ke 50 iterations se auto-optimize kiya. Outcomes: Iska result yeh nikla ke hamen 98.6% mAP@50 ki kamaal ki accuracy mili, aur GA ki wajah se hamara fitness score 81% barh kar 0.31 se 0.565 par pohnch gaya.

[Audio] Toh yahan hum dekh sakte hain ke Genetic Algorithm se pehle humara model kaisi state mein tha. Humne YOLOv8n ka pretrained model use kiya, batch size 16 rakhi, aur image size 640 by 640. Learning rate, momentum, weight decay — yeh sab default values hain, koi tuning nahi ki gayi abhi tak. Aur jab hum ne is configuration ke saath model train kiya, toh hamara initial fitness score — yani mAP50-95 — sirf 0.3120 aaya. Yeh hamari baseline hai. Ab hum dekhenge ke Genetic Algorithm in hyperparameters ko optimize karke is score ko kitna improve kar sakta hai..

[Audio] Ab is slide mai dekh skty hain — 50 Genetic Algorithm iterations ke baad hamare hyperparameters automatically optimize ho gaye. Learning rate thodi kam hui, momentum badh gaya, aur loss weights bhi adjust ho gaye — especially class loss jo 39% tak kam hui. Aur in sab changes ka result? Hamara fitness score 0.3120 se badh ke 0.5645 ho gaya — Yeh 81% ka improvement hai — sirf hyperparameters tune karne se, bina model architecture change kiye.

[Audio] Aur yeh hai humari final comparison slide — dono results ek saath. Left side pe dekhen — baseline model, jahan sab kuch default tha. Fitness sirf 0.3120. Right side pe — GA-optimized model. Learning rate kam, momentum zyada, loss weights refined. Aur fitness 0.5645 — yani 81% better. Sabse important cheez yeh hai ke humne model nahi badla, architecture nahi badli — sirf hyperparameters tune kiye Genetic Algorithm se — aur itna bada farq aa gaya. Yahi is research ka core conclusion hai.

Training Results — Loss & Metrics Over 50 Epochs.

Final Performance Metrics — Epoch 50. Summary of best validation metrics achieved after 50 training epochs.

Confusion Matrix — Normalized. Per-class prediction accuracy across all 13 behavior categories.

Precision-Recall Curve. Trade-off analysis — mAP@50 per class and overall.

F1-Confidence Curve. Optimal confidence threshold = 0.422 achieving F1 = 0.91.

Precision & Recall — Confidence Curves. How precision and recall change as confidence threshold varies.

Genetic Algorithm — Fitness Evolution Over 50 Iterations.

Hyperparameter Evolution — 50 Iterations. How each hyperparameter evolved during GA search (color = fitness score).

Hyperparameter Scatter Plots — GA Interaction Analysis.

Validation Predictions vs. Ground Truth. Model outputs compared against annotated validation labels.

Conclusions & Future Work. ✅ High Performance Achieved.