Abstract: To reduce the complexity of the driver detection algorithm and improve the accuracy of the seat belt recognition algorithm, a method based on convolutional neural network for driver detection and seat belt identification is proposed.By mitigating the cascading network framework, adjusting the feature training ratio, the driver candidate frame is generated as quickly as possible, and the correlation between the depth feature difference, the detection and the frame calibration is utilized to accurately locate the driver's position.By improving the classical convolution neural network, combining the maximum and average pooling layers, the total connection is reduced, and using batch processing of features to reduce the computational complexity and improve the accuracy of seatbelt recognition.The experimental results show that, compared to the other methods, the comprehensive evaluation criteria of the driver detection algorithm increased by 6.7%s, besides the accuracy rate of satety belt recognition increased by 3.4%. The algorithm meet the real-time requirements.