Object Detection Using a YOLOv4 Model in Video Surveillance

One of the most significant issues in surveillance videos is the detection of vehicles and objects at real time. Adverse weather conditions and lighting (illumination) variations are a few of the challenges faced while detecting objects through surveillance videos. In this work, a You Only Look Once v4 model, combined with a Spatial Pyramid Pooling and a Path Aggregation Network block and a contrast enhancement framework for effective detection of objects with higher accuracy is proposed. The video is converted to frames and is completely enhanced throughout using the Exposure Fusion Network. The brightness of the frames is equalized such that every point in the frame is clearly visible to the naked eye. Finally, the objects in the frames are detected using a YOLOv4 model, which is coupled with a SPP layer and a PAN network. The model generates an accuracy of 92.6%.