Measuring the Vehicle-in-Motion, Density and Allocation of Traffic Signal Using Transfer Learning

There are various deep-learning methods available for detecting and classifying vehicles. In addition to detection, a total count of vehicles helps build an intelligent traffic signal by allocating the time according to the vehicle density. But in real-world applications, the multi vehicles cannot be identified in the traffic signal, and the motion of vehicles identification rate is slow in real-time implementation through traffic signal-CCTV cameras. To overcome this, a transfer learning approach has been implemented to speed up detecting, classifying, and counting the vehicles in the traffic signal which helps to allocate time for traffic signal. There are many existing models which use different deep-learning techniques for the vehicle’s detection and classification. One such is YOLO V4, a recent super-fast model, one-stage object detector, and is best for real-time applications for object detection. YOLO V4 has been used for vehicle detection because of its very high inference speed when compared with other models. Having high inference speed is good advantage for real time test scenarios. Hence YOLO V4 has been used for implementation. The total estimation time is calculated using Transfer learning, where it’s based on the number of vehicles at each traffic signal. In this work, our model attains a mAP of 80% with 14–15 fps of speed with multiple vehicles in traffic signal.