Self-annotated Labelling and Training Data for Traffic Video Object Detection Using Machine Learning Techniques

Classical machine learning algorithms are susceptible to objective elements like video quality and the weather, which results in inferior detection results in an erroneous identification of Unmanned Aerial Vehicle (UAV) action photos. Deep learning is suggested for identifying cars in UAV video. Thanks to the technology offered by road management systems, real-time visual information is now accessible in thousands of locations on road networks. The first step in identifying or preventing accidents is locating the vehicles on the path of travel. Convolutional neural networks have made substantial advancements in the field of object detection. Enhancing conventional computer vision techniques. However, there are drawbacks because the pre-trained models that are currently available only provide a low detection rate, especially for small objects. The major drawback is that, in order to retrain the automobile maker, they must manually identify the vehicles that appear in the photographs captured by each IP camera on the road infrastructure. This task will be impossible to complete even if we install hundreds of cameras throughout the extensive road network. The technique utilized in this study uses CCTV footage with a comparable range of images with the aim of identifying cars in an image. The spatial brightness of the original sample is first translated using the HSV (Hue, Saturation, Value) method to maximize sample diversity and adaptability to different lighting conditions. The performance of the basic SSD is enhanced by incorporating focus loss because of feature extraction. Following a trained neural network model’s examination of the UAV footage, the effectiveness of drone identification is assessed. The results show that the proposed method have a car detection rate of 96.49%.