Abstract: To realize safe obstacle avoidance for unmanned aerial vehicles (UAVs) during low-altitude missions based on real-time obstacle perception, a constraint-driven end-to-end safe control framework is proposed based on a differentiable control barrier function. The method takes real-time sensing data, such as images acquired by the UAV, as input, utilizes neural networks to learn the relaxation parameters of safety constraints and mission constraints, and embeds them into a differentiable quadratic programming to generate control volumes. The gradient descent formula of the loss function for backpropagation is derived. A simulation scenario is constructed on the Airsim platform , and a process from training dataset construction to model testing and deployment is provided. The results show that the proposed method can generate control volumes end-to-end based on such perception data as depth images, etc. even when obstacle positions are unknown, achieving better optimal task success rates compared with those of fully connected network-based control prediction methods.