H-YoLov3: High performance object detection applied to assisted driving

For the application scenarios of assisted driving, a new detection model H-YoLov3 with higher accuracy, faster speed and smaller space occupation based on the YoLov3 is proposed. First, the FM module and the method of gradually increasing feature channel number with the perceptual field are proposed to redesign the feature extraction network. Second, constructing a neck of integrating target and environment using depthwise separable convolution. On the autonomous driving dataset BDD100K, the average accuracy of H-YoLov3 is 48.5%, the model size is 23.77MB, and the running speed is 35FPS. Compared with the YoLov3, the accuracy of the algorithm is improved by 8.4%, the model size is reduced to 10%, and the running speed of the model is also improved by 15FPS. This indicates that the H-YoLov3 provides better properties for the application scenarios of assisted driving and can be used in mobile or embedded scenarios.