The complex operating environment of the robot and the randomness of material distribution lead to low accuracy of robot target pose identification and positioning, and poor real-time performance. Therefore, a method for robot target pose recognition based on improved PSO-BP algorithm is proposed. The target image is preprocessed by an improved median filter algorithm, a multi-scale gray difference operator and a local image entropy operator are constructed, and the weighted local entropy is obtained by dot product operation to suppress the noise in the target image. The robot target pose features are extracted through the associated feature information of the multi-view geometric intermediate frames. In the BP neural network training stage, the improved PSO algorithm is optimized, and the optimized BP neural network algorithm is used to train and recognize the extracted features, and finally realize the robot target pose recognition. The experimental test results show that the luminance variance of the proposed method is 0.305 when the number of robot target test samples is 55, and the positional recognition error of the robot target obtained by the proposed method is 0.11 when the pixel recognition error is 1.5%. The proposed method can accurately recognize the robot target under the pixel recognition error and obtain the high-precision robot target positional recognition results.