| Abstract |
Video analytics is essential in smart city management, traffic monitoring, and security surveillance, where real-time decision-making is critical. However, the efficiency of these applications depends on optimizing video compression parameters to maintain high detection accuracy while minimizing bandwidth usage and computational costs. This paper presents a comprehensive survey of video compression optimization techniques, focusing on Quantization Parameter (QP), Frames per Second (FPS), Entropy Coding, and Motion Estimation for video analytics tasks. We examine traditional compression algorithms, machine learning-based approaches, dynamic parameter adjustment strategies, and hybrid models, each offering unique strengths and limitations. Our findings highlight that adaptive compression techniques improve the trade-off between detection accuracy and efficiency. However, challenges remain, particularly in dynamic and bandwidth-constrained environments. To evaluate these techniques, we employ AccMPEG. This video compression framework dynamically adjusts compression settings based on real-time inference feedback, ensuring an optimal balance between detection accuracy and bitrate efficiency. In addition, we assess perceptual quality using Just Noticeable Distortion (JND) and Mean Opinion Score (MOS). The results indicate higher QP values lead to noticeable quality degradation, particularly at lower bitrates. Motion estimation techniques influence perceived quality, and TESA achieves MOS ratings higher than alternative methods. Furthermore, H.265 demonstrates superior MOS scores compared to H.264 at lower bitrates, reinforcing its higher compression efficiency while preserving visual clarity. Despite these advancements, the generalization of findings is limited, as the dataset consists mainly of traffic videos, which may not fully represent other video analytics applications. Future research should explore a broader range of datasets and develop adaptive compression frameworks that integrate real-time inference feedback and advanced machine learning models. © 2013 IEEE. |
