Daniel Guo¹ and Aleksandr Smolin², ¹USA, ²Computer Science Department, California State Polytechnic University, USA

Computer vision models usually focus on trash detection in nature such as forests, oceans and beaches [1]. However, highway trash is a very common yet neglected problem, different from the trash found in nature, that many communities struggle with. When you drive along the highway, you can often see many concentrations of trash such as drink cans and plastic bags. This paper looks into different computer vision models to best discern the different varieties of trash found on the highway roads [2]. We compared two commonly used computer vision models, Yolov5 and Retinanet, to find out which one would best suit the applications of a highway trash detection model [3][4]. These computer vision models were trained on a kaggle dataset of commonly found trash and their categories; we first looked at the different validation steps of the models such as the environment of images and then we tested the two different models on urban images of a variety of trash, allowing us to determine which model was most fit for the discovery of trash on highways. Computer vision models have traditionally been applied to detect trash in natural environments such as forests, oceans, and beaches. However, highway trash is a widespread yet often overlooked problem that many communities struggle with. Unlike natural debris, highway trash is typically composed of drink cans and plastic bags, among other items. In this paper, we investigate the use of computer vision models to identify the different types of trash commonly found on highways. Specifically, we compare two widely used models, Yolov5 and Retinanet, to determine which one is better suited for developing a highway trash detection model [5]. We trained these models on a kaggle dataset that contains various categories of commonly found trash and validated the models under different conditions. Subsequently, we tested both models on urban images of a variety of trash to determine their performance in identifying trash on highways. Our results suggest that Yolov5 is better suited for this task.

Yolov5, Retinanet, taco trash dataset(Kaggle), Python