| Summary: | Children being trapped in vehicles is a critical issue that has led to many deaths. The number has increased dramatically regardless of the precautionary measures taken by their parents and guardians, which not only become a local but also a global concern. Though several technologies and systems have been developed to identify the in-car-abandoned children (iCAC), there are still reported cases occurring every day due to several factors. Deficiencies in the existing device need to be identified and improved to detect the children precisely. Over the past few years, the use of deep learning (DL) algorithms for remote sensing image analysis has risen, and several intelligent and efficient detection algorithms have been proposed. Yet, the improvement strategies DL-based object detection has never been summarized in detail. This paper will evaluate the accuracy of iCAC's images taken from the ground platform using DL algorithms. Besides, the hyperparameter is tuned to optimize detection accuracy. Convolutional Neural Networks (CNNs) and MobileNet V2 are two types of DL image classification models utilized in this study. Both models were used to extract features and train the three categories of child, no-child, and adult. Next, the classification outcomes are compared, and the model's effectiveness is evaluated. Learning rate, epoch and batch size were applied and tested to examine the change in the model's accuracy rate for the model's optimization. The results obtained demonstrated that the accuracy of children's image detection could be improved by optimizing the DL algorithms. This research contributes valuable insights into real-time image detection because the DL models have shown promising results in advancing the capabilities of image detection. The findings are significant to the ongoing optimization of remote sensing techniques for accurate and efficient image recognition. © Published under licence by IOP Publishing Ltd.
|
|---|