A robust traffic scene recognition algorithm based on deep learning and Markov localization

This paper designs a traffic scene recognition module for the agent's perception system. First, we enabled the output features of the convolutional neural network to be the descriptor of the traffic scene and adapted to the cost function of the image sequence to construct the observation module of the agent. Second, we assumed that the movement of the agent would be recursively updated and wouldn't jump dramatically, which simultaneously possesses the Markov property, so the Markov localization algorithm was used to improve overall robustness. Third, the Kalman filter method was adopted to represent the probability distribution of the entire system using the first and second moments of the Gaussian distribution, so that the loop iteration in the state estimation can be transformed into a linear operation, and the penalty term in the standard variance of the observation probability can also be added to describe the reliability of the observation. Experimental results show that the agent can efficiently remove unreliable observations and achieve robust recognition accuracy of the traffic scene in all weather conditions.