Nikhil CG

The number of fatal injuries on active construction sites has risen by nearly 90 percent in the last three years and our company is not immune. The four most common workplace hazards are falls, electrocutions, struck by and caught by accidents that constitute over 60 percent of such injuries. This project provides a solution that aims to improve worker awareness and company profits by reliably notifying workers of their proximity to potential hazards and active machinery. For our project we would like to develop a system that combines image processing and point-cloud based computer vision methods in order to track the movement of workers around the construction site and warn them of entering potentially dangerous regions. We would target the most common construction site hazards, namely: workers falling and being struck by objects. With this technology, our company will be able to easily identify the hazardous areas on our construction sites and prevent some of the most fatal accidents that we see every year.

For the design of the project, Our method is composed of 5 main steps. The first is to preprocess our worksite by selecting four fixed points in the cameras field of view. This allows us account for the unique perspective of our camera and generate the respective transformation matrix. Next, we read our video stream, and divide it into individual frames that You only look once (YOLO) can process. Applying YOLO allows us to identify workers in each image, and using the outputted bounding box, we can map worker positions using our previously defined transformation. Finally, we generate a map of worker positions for each frame and stitch that back together to form an output video.

Given a video stream of our worksite, we need a way to map worker positions in the image, to worker positions in space. To do this we rely on the fact that the camera’s position on the construction site is fixed and given 4 known positions we can generate a transformation that maps pixel values to world coordinate values. This works because we assume all of the workers are on the same plane, implying that their height is the same all times. So as you can see above, x, y are world coordinates and u and v are pixel coordinates.

After that we run YOLO on each of our frames, were given the pixels value for the center of the bounding box, and its dimensions. By averaging the bottom of our bounding box, we can find a pixel representation of each worker’s position in each frame. We transform those pixel values into world coordinates, using the matrix we found earlier. To generate our output video, we scale our world coordinate positions by the size our output frame and boundary. We then stitch together each frame to create an animation. With our predefined areas, we can accurately classify workers in danger areas.

The project was able to successfully track the location of the workers and plot their location on the bird-eyes view. We were able to track multiple bodies simultaneously and track their locations accurately. This was achieved with the use of just one camera.