Using Sensors to Complement Cameras to Save Arithmetic Power for Autonomous Driving Recognition

For the development of autonomous driving technology, object recognition is undoubtedly an important issue that cannot be ignored. Existing object recognition methods rely heavily on large deep neural networks, but there are still scenarios of incorrect recognition. Large deep networks also lead to an increase in recognition arithmetic. We work on an approach that addresses the scenarios where recognition errors still occur at this stage and does not require training giant networks. Our approach essentially explores the causes of recognition errors and analyzes the characteristics of the erroneous scenes. Other onboard sensors are then used to help the camera recognize the object. This avoids training deep neural networks to save computing power. We tested our approach on an existing dataset of bad weather and object occlusion and compared it with 2D detection and other sensor fusion methods. It can be seen that the mAP of our method increases by about 0.2 compared to the 2D detection, and the inference time saves about 60ms compared to the general 3D fusion method.