Semantic Segmentation with Deeplabv3 for Privacy Preservation in Intelligent Transport System

Intelligent transport systems (ITS) with camera imagery have made rapid progress. However, there is a threat of privacy-preservation violation, where the camera can capture human movement, which is not needed in ITS. This study proposes semantic segmentation for privacy preservation in ITS and evaluates its performance. To streamline the discussion, we refer to this approach as SS4PP: Semantic Segmentation for Privacy Preservation throughout the paper. We obtain image datasets for semantic segmentation from visual object classes (VOC) segmentation datasets. Deeplabv3, fully connected network (FCN), and lite reduced atrous spatial pyramid pooling (R-ASPP) are three semantic segmentation methods that are compared, where they are combined with three backbones, namely ResNet50, ResNet101, and MobileNetV3. We use Gaussian blur to blur the human, which semantic segmentation detects. Pixel accuracy and structural similarity (SSIM) index are chosen for testing metrics. The test results show that Deeplabv3 with MobileNetV3 backbone outperforms four other models that combine the mentioned semantic segmentation methods and the backbones. This is related to aspects with pixel accuracy of 0.92 and SSIM of 0.367. In addition, the test results also show that SSIM is more related to pixel accuracy than peak signal-to-noise ratio (PSNR) with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">${R}^{2}$</tex> of 0.2 compared to 0.1. Based on our SS4PP implementation on ITS, it is a challenge to create a more robust semantic segmentation of objects in cropped, small, and clustered images in the future.