Abstract. One of the objectives of Precision Livestock Farming (PLF) is to develop on-line tools for fully automatic and continuous monitoring of farm animals. PLF consists of measuring animal variables, modeling these data to select information and calculate specific parameters, and then apply these in real time for monitoring and controlling purposes. Computer vision is widely applied in PLF, as well as image analysis techniques that have been used to e.g., automatically detect lameness in dairy cows. Therefore, the performance of such computer vision systems relies on the quality of recorded videos. In particular, the aim of this work was to develop filtering procedures to exclude recorded dairy cow videos of low quality in order to optimize the performance of a lameness detection system based on images. An automatic filtering was developed to remove videos with: 1) Low 3D image quality; 2) Stopping cows or cows not walking smoothly. A top view 3D camera capable of retrieving depth information about cow’s body was used to record videos. Regarding the low 3D image quality, filtering was based on certain features extracted from the videos, such as Euler's number, spots ratio, roughness of cow’s body contour line and the number of spots around the spine line of a cow. A reference dataset with a total of 313 video instances was manually built up and labeled as low (223 instances) or high quality (90 instances). Furthermore, cows which stopped or did not walk smoothly during video recordings were filtered out from the videos dataset, based on extracted walking indexes from the images. In this case, an additional dataset of 276 videos with 82% normal walking cows and 18% of stopping cows was used as reference for the filtering validation. A decision tree algorithm J.48 was used for classification. The results led to two simplistic filtering algorithms based on a threshold of 0.4% spots ratio and a dimensionless walking index with value 3. The obtained classification accuracies were about 76% for the low 3D image quality filtering and 91% for cow’s walking speed filtering. These filtering results are considered satisfactory in terms of its application on an automated lameness detection system for dairy cows.