An Adaptive Vehicle Detection Algorithm for Traffic Applications

Deep learning-based vehicle detection methods have been widely employed in traffic applications. However, improving the performance of models in terms of both accuracy and efficiency is a challenging task. While a larger model can provide higher accuracy, a smaller model can be more efficient. In addition, since most deep learning models are trained on fixed domains, their performance may drop significantly in entirely different domains. In our work, a straightforward online adaptive framework for vehicle detection is proposed. In this framework, a large-scale vehicle detection model is adopted to produce benchmark labels and a small-scale inference model with higher efficiency is proposed. The adaptive property of the inference model is achieved by the proposed online training and distillation approach. Experiments demonstrate that the proposed approach can operate in real-time on embedded devices and maintain high detection accuracy in the face of changing scenarios.