The vision of ultra-low-altitude urban drone operations at high density is depicted by the introduction of concepts such as urban air mobility and the low-altitude economy, and a technological wave in high-density drone operations is generated. Collision avoidance technology is recognized as one of the key core technologies for ensuring the safety of high-density drone operations, and its impact on operational reliability is directly acknowledged. In multi-drone collision avoidance scenarios within localized high-density airspace, real-time and reliable collision avoidance strategies are found to be beyond the capabilities of existing decision-making methods. To address this issue, a high-density collision avoidance decision-making method based on the D3QN (dueling double DQN) algorithm was proposed. Specifically, the state space was reconstructed using fixed sectors, thereby enabling high-density collision avoidance inputs to be processed while neural network stability was ensured. Furthermore, existing civil aviation air traffic control collision avoidance rules were referenced, and a two-stage collision avoidance decision-making method was designed, which comprises both rule-guided and free collision avoidance components, based on homotopy theory. Simulation experiments have demonstrated that, compared to conventional methods, a collision avoidance success rate exceeding 99.66% is maintained in airspace densities greater than 30 flights per square kilometer, thus providing a feasible collision avoidance strategy for high-density urban drone operations.