DAFV: A Unified and Real-Time Framework of Joint Detection and Attributes Recognition for Fast Vehicles

In the past decade, with the development of computing equipment and CNN, target detection has made great progress, which has promoted the development of specific target detection. The purpose of vehicle detection is not only to extract the vehicle from a large number of traffic surveillance cameras, but also for some follow-up research, such as the structured storage of vehicle information, which needs to quickly identify the attributes of the vehicle. Based on those demands, we propose a method of joint Detection and Attributes recognition for Fast Vehicles (DAFV). Firstly, we present Feature Rapidly Extract Module (FREM), which is to quickly shrink the feature map size and enhance the runtime efficiency. Secondly, we present Feature Refinement Module (FRM) to increase feature utilization rate and improve the performance. Lastly, we present the Cross-Stage and Multi-Scale (CS-MS) Module to optimize scale-invariant design. Related experiments based on UA-DETRAC dataset proves that DAFV is a feasible and effective method. The DAFV is fast and the speed does not change with the number of vehicles. For 416 x 416 pictures, DAFV can reach 53 FPS with only 775 Mib GPU memory, which can meet the needs of real-time applications.