At present, the multi-targets autonomous vehicles identify are mainly various road users such as vehicles and pedestrians in cities. However, in recent years, there are increasing vehicle-animal accidents in pastoral areas of China, and there is a lack of corresponding practical intelligent safety warning measurements. In order to realise the identification and collision avoidance of typical animals, this study obtains the basic characteristics of collisions and the animals’ behavioural characteristics from the pastoral accidents investigation. Firstly, an attempt to focus on the use of the algorithm of the binocular infra-red equipment is made, with the main targets of identification being domestic animals at night or under low brightness conditions. Secondly, the algorithm technology roadmap from the identification and the risk assessment to collision avoidance decision-making is designed. Further, based on the technical route, a feasible solution and overall technical program is presented, which outlines the five aspects of the proposed engineering-ready scheme’s architecture. The methodology used involves transfer learning and the PSMNet network model. The paper concludes that the system can predict the animal’s movement and make proper decisions by utilising binocular infra-red equipment, integrating its algorithms and the support of its behavioural risk assessment. The research results will contribute to the animal collision avoidance technology and the traffic safety in pastoral areas in the future.