Computer Vision News - May 2021

236 Computer Vision Application Most of the common object detection datasets are ground perspective, so the team had to create their own, collecting around 5 terabytes of imagery from existing traffic cameras. They worked with another company to accelerate the data labeling process and then retrained everything. “On the tracking side we’re using a sort algorithm called DeepSORT, ” David tells us. “It’s a standard Kalman filter kind of approach, typical object tracking, but we are also using the appearance descriptor of DeepSORT to make sure that the unique ID of a car isn’t changing as it moves along. We already have the satellite imagery and street data, so we take the GPS location of every camera and perform a calibration that registers the camera’s view to a map. We’re then able to determine exactly where a vehicle is and can track its speed.” Speed + GPS + Road Name Tracking: showing same capability while pan/tilt/zoom operation is performed mid-way through video