Abstract: Abstract: Body weight is one of the main growth indices in broiler production, which is a comprehensive parameter in the broiler growth. The most common method to measure weight is manual operation, in which the broiler is captured and placed on the electronic scale. This method decreases animal's welfare and increases labor; in addition, it also will affect the yield and quality, and even cause the death of broilers. It can't be applied in commercial farms. The Kinect 3D (three-dimensional) camera which can measure the phenotype features with a non-invasive way has been applied into animal's weight acquisition. A broiler quality estimation method based on depth image was proposed in this paper. Yuncheng partridge shank chickens were chosen as research objects and an image collection system was constructed in a local farm. In this experiment, 150 broilers were selected randomly and the duration was the lifespan, 30 days. The acquisition system is composed of a Kinect depth camera, an industrial control computer, a serial port switching electronic scale and a fence. The procedure of image preprocess consists of image cropping, median filtering, OTSU threshold segmentation and binarization. And the maximum target in the binary image after morphological reconstruction, such as opening and closing, was recognized as object. In the feature extraction stage, 9 features were extracted using a mathematical geometry method, including area, eccentricity, width, length, radius, perimeter, volume, back width, and day age. In the model construction stage, a BP (back propagation) neural network was designed with 9 feature inputs and 1 weight output. After sampling randomly, 1985 samples were used as the training set, and the remaining 20% were used as the test set. Based on these data, the body mass estimation model was established to realize the population mass estimation. Compared with the measured results, the estimation has good performance. The root mean square error (RMSE) is 0.048 kg, the mean relative error (MRE) is 3.3%, the optimal fitness is about 0.994 3, the minimum relative error is 0.5% and the maximum relative error is about 11%. Different feature group and BP neural network were designed and trained. From the results of different modeling, it can be seen that the influence of 3D feature on the body mass is smaller than that of 2D feature. For the 3D features, the target volume has the least impact on the results, and for the feature from 2D group, the projection area has the greatest impact on the results. The fitting results of model which used 9 input parameters were the best. These results show that the proposed method is feasible and effective for constructing broiler quality estimation model. It provides theoretical basis for estimating broiler growth with machine vision technology as well as precision feeding.