Loading...
1 results
Search Results
Now showing 1 - 1 of 1
- Waste Container Detection System using Computer VisionPublication . Martins, João Miguel Baltazar; Neves, João Carlos RaposoWe propose a computer vision system for the automatic detection and counting of urban waste containers in video streams captured by garbage collection vehicles. Designed to support smart city infrastructure, the system enables geolocated container mapping and route optimization. Our approach is validated on a two-phase dataset comprising 144 videos ( 49 minutes) with over 35,000 annotated instances spanning 379 unique containers. We benchmark two detection models—YOLOv11 (image-based) and DiffusionVID (video-based)—across both dataset phases. YOLOv11 consistently outperforms DiffusionVID, particularly on the augmented dataset, achieving a mAP@0.5 of 0.938, despite the latter’s strengths in detecting small-scale objects. For counting, YOLOv11 is integrated with ByteTrack and enhanced using three domain-specific heuristics: (H1) short track filtering, (H2) identity merging, and (H3) spatial consistency. This configuration yields substantial improvements in accuracy, reducing the Mean Absolute Error (MAE) and Sum of Absolute Differences (SAD) by up to 77% on the augmented dataset. System robustness is further validated on real-world deployment videos ( 2 hours each), demonstrating that the effectiveness of heuristics varies from video to video. Nonetheless, the H1+H2+H3 combination demonstrates the best generalization and is recommended for practical deployment. Our contributions include: (i) a novel annotated dataset for urban waste container detection, (ii) a detection–tracking pipeline, and (iii) tailored heuristics for improving counting accuracy. Future work will address class imbalance, conduct failure case analysis, and evaluate scalability on continuous, long-duration video streams representing full waste collection routes.