As the complexity of transportation systems increases, driving safety has become a critical issue in these years. Object detection based on Convolution Neural Networks (CNNs) plays an important role for intelligent driving in detecting vehicles, pedestrians and so on. However, the detection models in CNNs ask for huge computation and memory usage, which makes it difficult to work in on-board devices. What's more, intelligent driving needs high-precision and low-latency requirements in city traffic scenarios for object detection. This paper proposes an intelligent cooperative vision perception system for the connected vehicles. The driving videos are collected from vehicles and transmitted to cloud for vision perception with CNNs in this system. In addition, there are two kinds of visual models which achieve 73.64% and 65.67% mean Average Precision (mAP) respectively for chosen. Considering the bandwidth consumption and transmission latency, a resolution adaptive strategy and a frame-rate control strategy are designed and the IPv6 route is used for transmitting between the vehicle and cloud. This system assists drivers for better and safer decisions by augmenting the perception ability.