In this study, a bottle counting system is developed based on processing of depth images. 3D depth images are captured by Kinect-2 sensor, which was mounted at a facility recycling line. Based on the facility's demand, three different types of bottles were considered in the bottle counting system. Boxes contain either 20 or 30 bottles depending on the bottle type. The facility recycling line moves at 1 m/s velocity and at least one box crosses from facility line per second.. The bottle counting system consists of an image processing algorithm for processing captured depth images. The first step of the algorithm includes camera calibration routine so as to convert sensor measurements from camera frame to a fixed world frame. Once the measurement is triggered, morphological operations are applied to the binary image to remove noise and close the gaps. Connected component analysis, min bounding rectangle is performed to detect objects in the image. The cycle time for the developed image processing scheme is 45 ms: 33 ms for image capturing (30 fps), 5 ms for the pre-processing and 7ms for the bottle detection steps. It is observed from the results that the developed algorithm can count the bottles with 99% accuracy for each bottle type.