The transformed image was filtered by convolving with the Mexican hat wavelets, followed by a search for local maxima corresponding to the position of individual crop plants. For ease of computation the convolution was conducted locally based on initial predicted plant positions from a Kalman filter tracking algorithm (Bar-Shalom & Fortmann, 1988). This algorithm was based on earlier work described by Hague and Tillett (2001) in which the two most important Kalman filter states were heading and lateral offset with respect to crop rows. In this work these states are augmented with additional states allocated to individual plants representing forward distance ahead of the toolbar. As the number of plants in a scene varies with time the state vector was dynamic with new states created as plants appeared at the top of an image and deleted as they pass the cultivator. Having used the Kalman filter estimate for initial placement an adaptive step size hill climbing technique was used to accurately position a Mexican hat over individual plants. Local maxima representing plant centroids were then used to refine the Kalman filter estimate of plant positions.