Monocular depth estimation aims to predict depth information within a scene from a single RGB image, but many models remain computationally intensive for real-time inference on resource-constrained edge devices. This paper presents a lightweight self-supervised monocular depth estimation network that balances accuracy and efficiency through targeted encoder–decoder design. The encoder employed a synergistic modeling approach combining decomposable large-kernel convolutions and local depthwise convolutions to capture both long-range context and local details with low computational overhead. The decoder utilized cross-scale feature differences as guidance to dynamically fuse multi-scale features, enhancing detail recovery and geometric consistency under lightweight constraints. In addition, a temporal soft fusion reprojection loss was employed to better leverage the complementary information of forward and backward frames, improving the robustness of self-supervised training. The model contained 3.0 M parameters and required 3.5 GFLOPs of computation. On KITTI, it achieves Abs Rel=0.105 and δ₁=0.892. On Make3D, it achieves Abs Rel=0.308 in a zero-shot setting. On a Rockchip RK3588S, a hybrid-quantized multi-thread implementation runs at 67 frames/s. The results demonstrated that the proposed method achieved a favorable accuracy–efficiency balance on edge devices, making it suitable for real-time monocular depth estimation tasks.

Aiming at the problems of difficult deployment and access of surveillance system server, as well as high operation and maintenance cost, a remote surveillance camera is designed based on RK3566 chip, which is controlled and transmits data via email platform. Firstly, to address the impact of environmental factors such as weather and light on image quality, a deep neural network (DNN) image exposure correction network is employed to rectify images with abnormal exposure. Additionally, a back propagation (BP) neural network is utilized to fit a curve relating the brightness difference to the gamma value of images before and after exposure correction, thereby adjusting the gamma value of the camera. Secondly, to enhance the precision of YOLOv5 algorithm in differentiating between anomalies in nighttime imagery, infrared image data are employed, and a context-aware light-weight label assignment head and coordinate attention mechanism are incorporated into the model to augment the model’s detection accuracy and recall rate for small targets. Furthermore, to meet the demand for reporting of abnormal situations in unattended environments, an automatic target identification and reporting process has been designed which combines YOLOv5 algorithm with the frame-difference motion detection algorithm. The camera has been tested for compatibility with the current mainstream commercial email platforms. The mean time required for transmitting a single image file via the email platform is less than 10 s, while the mean time for transmitting a short video is less than 60 s. The BP network’s average training loss is 0.015, and the average testing loss is 0.013, which basically meets the precision requirements for gamma adjustment. The improved YOLOv5 algorithm achieved an mAP@0.5 of 91.5% and a recall rate of 85.5%, effectively enhancing the accuracy of small object detection.