Abstract: Aiming at the collision avoidance problem of logistics unmanned aerial vehicle in the complex urban low-altitude environment and high-density dynamic traffic flow, a dynamic three-dimensional (3D) collision avoidance algorithm is proposed. Firstly, we model the urban low altitude operating environment, express the dynamic collision avoidance problem of unmanned aerial vehicle as a Markov decision process, and expand the feasible solution range of algorithm to the 3D space by adding the altitude change and other manoeuvre into the action set of collision avoidance. Secondly, we improve the reward valuation function, so that the algorithm can balance the optimal decision by Monte Carlo tree search in two-dimensional plane flying around and 3D space obstacle crossing. Finally, the global optimal solution is gradually obtained by approaching the single optimal feasible solution. The simulation results show that the algorithm can optimize the collision avoidance action, and choose the best collision avoidance strategy in flying around and crossing obstacles to shorten the flight time by 24.4% and reduce the flight distance by 33.2%. For the unmanned aerial vehicle operating in the urban low-altitude environment, its sense system is easy to be partially observable due to the limited building occlusion, and the algorithm cannot obtain sufficient environmental state information input for solution calculation, so the algorithm requires robustness. The simulation results show that with the shortening of sense radius, the algorithm has good performance and can still give most of the feasible solutions under the limited conditions of the unmanned aerial vehicle sense system.