Improving the comfort of vehicle braking is of great significance to improve the travel experience of drivers and passengers, but the problem is that compared with the dynamic calculation of conventional braking, the calculation difficulty is greatly increased, and at the same time, it is often at the cost of increasing the braking distance and reducing braking safety. In this paper, an intention recognition system based on multi-layer information superposition and a comfortable braking method that works together with the acceleration curve at the end of smooth braking are proposed to further improve the braking comfort of the vehicle and ensure the braking safety. Based on the LSTM neural network trained by the HO algorithm to predict the characteristic values of the braking intention determined by experiments, the driver′s braking intention is accurately identified, and the qualitative characterization is emergency braking or conventional braking. According to the comfortable braking expectation proposed in this paper, the comfortable braking expectation is quantitatively accumulated for all the above information, the comfortable braking expectation caused by safety and traffic efficiency is distinguished by 0 and 1, and the accumulation caused by eigenvalues is accumulated by the quantitative accumulation formula proposed in this paper, and the comfortable braking behavior is stimulated after reaching the threshold: the intention recognition system of multi-layer information superposition identifies the situation that is safe, does not affect the traffic efficiency and conventional braking intention, and smooths the acceleration curve at the end of braking. Based on the Carsim-Simlink platform, three working conditions were built for simulation and experimental verification. The results show that the indicators of comfort and safety are significantly improved, indicating that the method proposed in this paper has an improvement effect on the comfortable braking of automobiles.