Real-time Litter Recognition Using Improved YOLOv4 Tiny Algorithm

• Published in: 2022 IEEE 2nd Mysore Sub Section International Conference (MysuruCon) pp. 1 - 5
• Main Authors: V, Shalini, Tangade, Shrikant, K, Prajna P, J P, Sangeetha, Azam, Farooque, L, Anoop G
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.10.2022
• Subjects: fully connected layer; litter; Machine learning algorithms; Object detection; Prediction algorithms; prediction time; Predictive models; Real-time systems; Roads; Sociology; YOLOv4-Tiny
• DOI: 10.1109/MysuruCon55714.2022.9972356

Littered roads have become a familiar sight in India. The main reason is the increasing population and inefficient waste disposal system. Since garbage collectors cannot pick litter in all the places, there is a need for an efficient way to detect it. Hence, a machine learning-based object detection model is used. In this, we have applied an improved YOLOv4-Tiny algorithm to detect the garbage, classify it and make the detection process easier on custom datasets. We have improved the algorithm in terms of the object prediction time, this is done by replacing a max pooling layer with one of two layers present in a fully connected layer. When an input is given, the algorithm detects the litter in the image with a bounding box around it along with the label and confidence score. The proposed model reduces the prediction time by 0.517 milliseconds less than the original algorithm employed which concludes that the object is predicted faster.