Abstract: To address the problem that the traditional pedestrian detection algorithm extracts a considerable amount of redundant information and has low recognition accuracy, we propose a pedestrian detection algorithm based on specific area decorrelation. On the basis of the response value of the average shape template feature map, the head-shoulder region is the most discriminative region in the classification stage. By incorporating the priori statistical human information into the decorrelation operation of the feature channel, a new filter is generated by extracting the covariance from the head-shoulder region, which is shared with the entire detection window in each channel during the test phase. At the same time, the problem of cross-channel sharing of covariance is investigated. A large number of experimental results show that the proposed method can effectively reduce the average logarithmic missed detection rate while maintaining the same detection speed, thus meeting the requirements for real-time pedestrian detection.