Classification of Trash and Valuables with Machine Vision in Shared Cars

This study focused on the possibility of implementing a vision-based architecture to monitor and detect presence trash or valuables in shared cars. The system was introduced take pictures rear seating area four-door passenger car. Image capture performed with stationary wide-angled camera unit, image classification conducted prediction model remote server. For classification, convolutional neural network (CNN) form fine-tuned VGG16 developed. CNN yielded an accuracy 91.43% batch 140 test images. To determine correlation among predictions, confusion matrix used, addition, for each predicted image, certainty distinct output classes examined. execution time system, from capturing displaying results, ranged 5.7 17.2 s. Misclassifications were observed results primarily due variation ambient light levels shadows within images, which resulted target items lacking contrast their neighbouring background. Developments pertaining modularity unit expanding dataset training images are suggested potential future research.