AI + IoT Technology for Building an Intelligent Sanitation Ecosystem" Project Development

The "AI + IoT Technology for Building an Intelligent Sanitation Ecosystem" project aims to create a comprehensive public sanitation service system that operates as a closed-loop for "service, management, and disposal." This is achieved through the integration of intelligent terminal modules, an Intelligent Identification Water Control System, a Rubbish Identification Control System, and an IoT-based Smart Sanitation Cloud Platform.

Intelligent Visual System and Method

The system utilizes a central data processor and wide-angle vision sensors to extract information from images captured by these sensors, identifying whether rubbish is present in the images. It classifies the rubbish and controls an automatic sweeper to clean and categorize the rubbish, while also monitoring the cleanliness level in real-time. The system achieves a frame rate (fps) of 32fps for rubbish detection within a 15-meter range, with an Average Precision (AP) of 76.3%, a Precision of 76.7%, and a Recall of 89.2%.

Intelligent Watering Control System

This system employs precise computer vision detection and AI recognition algorithms. Using a central data processor and wide-angle vision sensors to capture ground pedestrian images, it extracts information from the images to identify the presence of pedestrians, while ignoring vehicles and other obstacles. If pedestrians are detected within the watering range, the system automatically shuts off the valve and resumes operation after avoiding the pedestrians. The system achieves a frame rate (fps) of 37fps for pedestrian detection within a 15-meter range, with an Average Precision (AP) of 74.5%, a Precision of 74.5%, and a Recall of 90.5%.

Jinkai Intelligent Urban Road Sanitation Monitoring Platform

The cloud platform developed in this project is optimized in terms of system architecture selection, high concurrency technology, wide compatibility, strong resilience, and high scalability. It efficiently processes data collected from terminal controllers, application scenarios, and command centers, forming a big data information management platform. The platform standardizes, analyzes, and utilizes the data to optimize operational routes, reduce operational costs, and enhance management efficiency.

Deployment of AI Deep Learning Algorithms on Low-Power Embedded Terminals

Traditional deep learning algorithms rely on extensive convolutional, pooling, and concatenation operations, resulting in high model complexity and significant GPU core usage, making deployment on low-power embedded terminals challenging. Jinkai's self-developed detection model uses ResNet as the backbone, replacing traditional convolutional kernels with dilated convolutional kernels and Bottleneck Layers to obtain larger receptive field feature maps. These feature maps are used as the output layer, with regression predicting the target image coordinates and sizes. Regularization terms are introduced to sparsify network weights, and the Hessian matrix of loss-weighted parameters is used to calculate parameter importance. Less important parameters and network layers are pruned, and the model is retrained and iteratively pruned. The pruned model parameters are then quantized to int8. The final pruned model retains only 47% of the original model's parameters, with accuracy and precision reduced by less than 3%, and inference speed increased by 60%, enabling efficient deployment on low-power embedded terminals.

Optimized Cloud Platform Technology

The system is built on the spring-boot and vue.js 3.0 platform architecture, employing the advanced Netty framework. Under single-point service deployment, the platform supports over 40,000 TPS (Transactions Per Second), with parallel support for hundreds of devices requiring long connections. The platform's data transmission efficiency has been optimized based on the MQTT protocol, ensuring high stability between the platform and device data transmission. The deployment solution uses nginx load balancing as a reverse proxy, effectively mitigating service disruptions caused by single-point failures. The platform uses a distributed file system (HDFS) with self-developed optimization techniques to enhance data throughput. Additionally, redis is used for local data caching, reducing database requests and ensuring database stability. The platform's microservice architecture allows for rapid response to customized business needs of different customers.