A study on urban street garbage detection based onYOLOX

Abstract Urban street garbage detection is a pressing issue for governments at all levels. The traditional manual detection method isbecoming inadequate to keep up with the demands of urban expansion. We must implement a more innovative approach toreplace it. Swift progress in object identification technology has provided a new method for identifying litter on city streets. Theexisting street waste detection is hindered by inadequate datasets and visual features that do not align with the semanteme.There is still a significant gap to cover in utilizing target detection technology for street litter detection. As a result, weproposed a new approach to tackle deficiencies in the dataset for detecting waste in urban streets. It includes a technique forexpanding data that can assist in resolving the problem of an incomplete dataset. In addition, we have introduced a ContextualTransformer (CoT) module based on YOLOX-S, which enables us to harmonize the disparity between visual characteristicsand semantics. By enhancing the aforementioned techniques, our approach achieves a 9% increase in detection accuracycompared to the original YOLOX-S on the identical dataset. Ultimately, it attains a precision of 92.82% in the mean AveragePrecision (mAP) statistic.