In this work we report an intra-fractional markerless algorithm that accurately detects lung tumors on mV projections within the beam's eye view, while minimizing harmful effects such as poor soft tissue resolution, global image distortion, image blurring and scattering due to intrafraction target motion and radiation scatter.. First, we generate two sets of DRRs digitally reconstructed radiographs-background DRR without tumor and tumor only DRR from the 4D CT planning data after the tumor has been initially segmented out. Next, the composite DRR is generated by fusing the tumor DRR on the background. The composite DRR along with the matching mV projection are divided into a matrix of small tiles. The tile configuration is automatically set up such that the tumor always remains within the beam's-eye-view geometry on the composite DRR. In order to locate the tumor on the mV projection, the tumor DRR is fused at different locations on the background DRR while the tiles of the composite DRR are globally shifted. For each configuration, the composite DRR is matched with the corresponding mV projection. A simple NCC normalized cross correlation is used to compute the similarity between the composite DRR and corresponding mV projection tiles. Finally, the location of the lung tumor on the mV projection is identified based on the best match found. The algorithm was successfully tested on a dynamic chest phantom at our institution. Approximately 5700 raw images over 12 gantry angles were tested and the tumor was accurately located on every mV projection. Although, the chest phantom was created to mimic the human chest anatomy with neighboring organs, tissues and bony structures, which introduced strong signals, the maximum error reported was less that 1.6 mm while the average error reported was less than 0.7 mm.