Developing a machine learning-based surveyor solution for a rapidly growing civil engineering consultancy

The Daffodil team tapped into its vast reserve of skills and know-how around AI domains such as AI image processing, analytics, and training ML models for industry-specific applications. They created a wireframe for an ML-based image recognition system to fit aptly into the vision and requirements brought to them by the civil engineering consultancy.

The system was built using a deep learning API known as Keras, which leverages the tools provided by the TensorFlow ML modeling platform. The final system created by them consisted of the following capabilities and modules:

The Algorithm

Earlier, a human surveyor had to label each element on the road in real-time based on the camera feed from the car mounts. Daffodil rehashed the whole system by eliminating the need for human intervention completely through an original ML-based algorithm. An ML-based computer vision technique was employed to identify instances of objects in the images and videos captured.

A variety of categories for the objects were defined including high tension wires, guard rails, beams, barriers, etc. The ML algorithm has been trained by feeding it 1000 images for each category of elements so that the survey report generation can be accurate and quick at the same time.

Continuous Improvement In Accuracy

An average of 87,000 image frames are generated per video recording, out of which 15,000 images are distinguished as relevant, finally filtering down to 1500 elements. This is after false positives, false negatives, and visibility limitations are factored in. This entire filtering process, despite its complexity, is completed in a matter of seconds with the new ML-based surveyor.

Location Calibrations

The camera captures the latitude, longitude, and altitude in the frame in video mode. Each image frame is separately analyzed by the ML algorithm converting the image to text and recording the location details as well. Additionally “chainage” is identified automatically, which means that the algorithm identifies where a long-running road element such as a barrier beam starts and stops. This helps differentiate between long-running elements and point objects such as signages and culverts.

Predictive Analysis

To further expedite the process of creating a road survey report, the ML system comes with predictive analysis capabilities also. A particular threshold has been prepared for recording probabilities for object detection. Based on these probabilities and advanced mathematical logic, the high probability objects appearing on a road being surveyed are recorded much faster through multiprocessing.

Leveraging Existing Infrastructure

There was a large volume of video and image data on the local servers of the consultancy which had to be run through the new ML model for training it. So the newly implemented system was exported and installed into the existing infrastructure of the consultancy which led to an immense reduction in the anticipated overhead.