Video Compression Effect on a Camera-based Occupancy Prediction Model

Abstract

Camera-based occupancy prediction models estimate the presence and distance of obstacles using camera images and camera positional data. This thesis investigates how lossy video compression impacts the performance of such a model. To test this, video compression was applied using three video codecs: MJPEG, H.264, and H.265. The compressed frames were then used to train 15 versions of an occupancy prediction model, and the resulting validation loss was analyzed for each configuration. The findings show that the selected occupancy prediction model demonstrated a high degree of resilience to compression artefacts. Although high-compression variants produced less stable predictions across consecutive frames, the predictions still contained mostly valid structures that resembled the expected outcome. As this study was conducted using synthetic data generated in the CARLA simulation environment, future research should explore whether models trained on real-world datasets exhibit similar robustness.