Vehicle Speed Estimation via Blob

Question

I have detected vehicles as a blob in OpenCV. Below is the blob.h file

class Blob {
public:
    // member variables 
    std::vector<cv::Point> currentContour;
    cv::Rect currentBoundingRect;
    std::vector<cv::Point> centerPositions;
    double dblCurrentDiagonalSize;
    double dblCurrentAspectRatio;
    bool blnCurrentMatchFoundOrNewBlob;
    bool blnStillBeingTracked;
    int intNumOfConsecutiveFramesWithoutAMatch;
    cv::Point predictedNextPosition;
    // function prototypes 
    Blob(std::vector<cv::Point> _contour);
    void predictNextPosition(void);
};

What algorithm should I use to estimate the speed of the detected vehicle??

Thanks in Advance.

UPDATE

Here is the code I have tried to estimate the speed, but it doesn't put the text plus it crashes.

for (auto blob : blobs) {

    if (blob.blnStillBeingTracked == true && blob.centerPositions.size() >= 2) {
        int prevFrameIndex = (int)blob.centerPositions.size() - 2;
        int currFrameIndex = (int)blob.centerPositions.size() - 1;
        if (blob.centerPositions[prevFrameIndex].y > (intHorizontalLinePosition-50) && blob.centerPositions[currFrameIndex].y <= intHorizontalLinePosition) {
            int distance = blob.centerPositions[currFrameIndex].y - blob.centerPositions[0].y;
            int tickCount = cv::getTickCount();
            int time = (tickCount - blob.firstTickCount)/cv::getTickFrequency();
            int speed = distance/time;
            double dblFontScale = blobs[currFrameIndex].dblCurrentDiagonalSize / 10.0;
            int intFontThickness = (int)std::round(dblFontScale * 1.0);
            std::cout<<"Speed: "<<speed<<std::endl;
            cv::putText(img, std::to_string(speed), blobs[currFrameIndex].centerPositions.back(), CV_FONT_HERSHEY_SIMPLEX, dblFontScale, SCALAR_GREEN, intFontThickness);
        }

    }
}

Answer 1

In order to predict the vehicle's speed in a 3-dimensional space from a 2D image in the general case, you need to know the orientation of the vehicle (direction of travel) and distance from the camera.

If you know for example that the vehicle is travelling perpendicular to the direction the camera points (moving directly across the frame, not toward or away from the camera at all), you can use either

a) A known distance from the camera to the road and basic trigonometry, or b) Markers of known distance

to calculate the velocity of the vehicle using several frames.

If you know the vehicle is travelling directly toward or directly away from the camera, you can use the change in width/height of the image outline to get a sense of the vehicle's speed. If you can also identify when the vehicle passes a landmark at a known distance from the camera, you can calculate the actual width/height of the vehicle and therefore accurately calculate the speed using that known width/height and rate of change of the size of the 2D projection of the vehicle.

Update

Given the additional information, it seems you can determine what Y position in the camera's 2D image corresponds to a particular distance down the road. If you measure two such points, you can count how long it takes for the lower bounds of currentBoundingRect to pass from the first point to the second point, e.g. in the diagram below to move from y=800 to y=200.

If it takes 2 seconds to move from y=800 to y=200, it also takes 2 seconds to move 100m - 50m = 50m, or 50m/2 seconds = 25m/second.