Self-Supervised Monocular Depth Estimation Using Temporal Convolution

In recent years, self-supervised monocular depth estimation has garnered significant research interest due to its lack of reliance on labeled data, adaptability, low cost, ability to combine with pose estimation, and continuous learning capability. However, existing self-supervised monocular depth estimation methods often produce ambiguous depth predictions for moving targets in dynamic scenes by due to inadequate modeling of temporal coherence. Therefore, this paper proposes an improved architecture based on Lite-Mono. We introduce a channel-space attention mechanism into the continuous expansion convolution module to enhance the fea-ture response to the edges of dynamic objects and untextured regions, thereby reducing depth blurring. A multi-scale tem-poral feature buffer is designed to store the multiscale feature maps of consecutive frames in a sliding window and generate timing-enhanced features through cross-frame fusion. A hier-archical temporal convolution architecture is proposed, stacking causal convolutional layers with an increasing dilation rate, to aggregate long and short-term temporal contexts and cap-ture the continuity of motion trajectories of dynamic objects. The decoder fuses temporal features with original spatial fea-tures to generate motion-consistent multiscale depth maps. Through extensive experiments, we compare the proposed method with the benchmark model Lite-Mono. Quantitative evaluations demonstrate superior accuracy of the proposed method and demonstrates superior performance in image edge and detail regions.