A real-time detection method of multi-scale traffic signs based on dynamic pruning strategy

Traffic sign detection can provide important judgment information for the unmanned driving system. To deploy the detector on edge equipment and have better detection performance, a real-time detection method based on a dynamic pruning strategy is proposed. First of all, we propose to add a shallow prediction branch to the initial model, which makes the three-scale prediction branch become the four-scale prediction branch. The preliminary improved network expands the detection range and improves the detection accuracy of the model. Secondly, we propose a dynamic channel pruning method to prune the model, which can greatly reduce the amount of calculation without reducing the accuracy of the model, thus improving the speed of traffic sign detection. Thirdly, the weight sharing method is used to replace different weights in the same cluster with the same value, which reduces the occupied hardware resources and realizes rapid deployment of the model. Finally, to increase the number of rare traffic signs, we propose to merge the TT 100 K and CCTSDB into one dataset. The experimental results show that the improved model not only improves the detection accuracy, but also improves the inference time, which is conducive to the deployment of the model on resource-constrained devices.