Advanced Solutions
With 2,615 km of sewer pipes, 3,906 km of water pipes and 885 km of stormwater infrastructure, Mangaung Metropolitan Municipality appointed SMEC South Africa to develop operational and maintenance plans for its ageing water and sanitation infrastructure. As part of the Asset Investigation, SMEC devised a Computer Vision Model to speed up analysis of underground CCTV footage.

Computer Vision Model

SMEC has developed an innovative solution to incorporate Artificial Intelligence (AI) into its workflow processes. Instead of reviewing thousands of hours of CCTV footage manually, the team is training several computer vision models for different types of infrastructure analysis to automatically extract relevant details from footage (video and images).


Following the success of these completed models, we are in the process of training a model to automatically identify the locality and severity of irregularities in stormwater and sewerage pipes from previous project data. This approach reduces human error and optimises time allocation, allowing our team to focus on more complex tasks.

Figure 1 displays an example of the results from the preliminary model, with the output generated through the following process:

Deploying Artificial Intelligence for underground asset condition assessments

  1. The in-office user selects the CCTV file or folder directory of multiple files and initiates the model.
  2. The model then runs without human supervision through each video frame and searches for any irregularities learned during the training stage.
  3. If the model identifies an object (irregularity), the model outputs the following data subject to a threshold confidence level:
    • If the confidence of the detection is high, the model outputs the data shown in Table 1. In addition to this output, the model can retrieve any metadata from the video file, as requested by the client. For example, if the drone is GPS enabled, then the X and Y GPS coordinates of the identified irregularity can be extracted.
    • If the confidence of the detection is low, the irregularity is identified as an edge case and the model trims the video for later inspection by the in-office technician.
  1. The model then pushes the extracted data to a database, or the like, for record keeping.
  2. Once the model is complete, the in-office technician needs only review the CCTV footage of objects identified by the model with low confidence (edge cases), thereby substantially reducing the time required for the review. As the model is used, the reviewed low-confidence detections are used by the SMEC AI team to improve the accuracy of the model.
  3. Thereafter, the Project Engineer can assess the structured data to make data-driven decisions on the pipe replacement or point repair requirements.