WasteDet: An Anchor Free Novel Detection Algorithm for improving Waste Management
Rishabh Tiwari, Ashwani Kumar Dubey · 2022 International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES) · 2022
Noteworthy steps are being taken towards improving the level of hygiene and cleanliness in cities. Even with these efforts littering and non-recyclable waste is still a serious major issue socially and ecologically. Locating, identifying, and picking up the waste by staff can be a tiresome and inefficient task requiring long hours of manpower. To address the problem of locating and identification of the waste objects we propose a cost-effective method which is based on application detection algorithms on low altitude imagery from streets, sidewalks, houses, and landfills taken by surveillance camera or Unmanned aerial vehicles or UAVs. The novel deep learning-based framework called WasteDet consists of deep convolutional neural networks which perform the localization and the classification task which can then be processed and put through into automated pickup planning system. A custom dataset was compiled and annotated to train and evaluate the model. The model achieved a mAP score of 87.82% when tested on 8 classes while testing on all classes a mAP score of 85.60% was obtained, outperforming the current state-of-art object detection algorithm by 4% AP. This model is also capable of detecting objects in real-time applications on live and recorded videos achieving 49.6 fps on our video dataset.