<p>This paper presents an adaptive approximate computing-based traffic sign recognition unit for self-driving cars called AdApTS. Traffic sign recognition is one of the critical activities of self-driving vehicles that should be performed accurately on the cameras’ outputs. To this aim, deep learning-based image processing schemes provide acceptable accuracy, but their computation overhead is very high for the resource-constrained systems. The proposed AdApTS employs non-uniform approximations in the computations of a convolutional neural network as the best image processing option. To this aim, the redundant and unimportant computations in kernel and input image processing are derived and approximated. The important information about traffic signs is generally embedded in their shape and center; thus, the dynamic segmentation of them leads to appropriate results. This significantly decreases the computation cost and response time while maintaining an appropriate level of accuracy. To demonstrate the effectiveness of the proposed AdApTS, several experiments on real-world traffic signs are performed. Based on these experiments, AdApTS improves the response time by about 71% in cost of 2.1% accuracy degradation by focusing on approximate kernel or input computations. In the case of combining the kernel and input approximation, AdApTS provides 92.8% accuracy in 2.43&#xa0;s. Compared to related approximation-based studies, our proposed AdApTS outperforms accuracy and response time by 13.03% and 14.8%, respectively.</p>