Due to the high computational costs of Deep Neural Networks (DNNs), leveraging edge servers for video analytics on mobile devices equipped with such applications serves as an effective solution. However, existing schemes fail to adequately consider the semantic correlation between image content and video analytics, resulting in high end-to-end latency on bandwidth-limited devices. In this paper, we introduce FRVA, a content-aware video frame redundancy elimination framework designed for bandwidth-limited edge video analytics. FRVA employs a lightweight Convolutional Neural Network (CNN) to perceive inference-relevant content within video frames, enabling content-aware frame filtering and ROI encoding. This approach filters out non-essential frames and removes non-target regions from retained frames. To maximize FRVA’s performance, an optimization problem is formulated to balance inference accuracy and end-to-end latency. Then, a Deep Q-Network (DQN)-based online configuration method is designed to adaptively select parameter configurations under dynamic system conditions. Consequently, FRVA can eliminate redundancies in video frames data significantly. Extensive evaluations demonstrate that FRVA achieves superior compression ratios and significantly lower end-to-end latency compared to state-of-the-art redundancy elimination methods.