In the multi-scale collaborative human head detection system, the gradient direction histogram cannot meet the real-time requirement of video surveillance because of the massive computation in the high-definition video surveillance field. This paper proposes a method of human head detection based on GPU_CPU heterogeneous parallel acceleration. The GPU is responsible for the HOG feature extraction of large-intensive block parallel computing, and CPU is responsible for the implementation of other modules. The traditional parallel reduction algorithm is not excellent in the HOG feature extraction, and an improved parallel reduction algorithm is therefore proposed, which reduces the time complexity by the parallel computing of down-sweep to reduce calculated times of nodes, and the experimental results show that the proposed method is more efficient than the traditional one for over about 10 times.