Object detection algorithm in video surveillance systems based on neural networks

The application of video surveillance system in modern society has become more and more common, however, they are still faced with many technical problems in accurately identifying the target object in the surveillance screen. As computer vision (CV) has evolved, neural networks have achieved tremendous growth within the domain of object detection. Aiming at this problem, this paper selects UA-DETRAC dataset as a reference, and takes vehicle type recognition as an example to deeply study the video surveillance system based on neural network. The system we built is based on the basic framework of Faster R-CNN (Faster Region Convolutional Neural Network) and introduces a series of innovative techniques to improve detection accuracy. First, we use color dithering and image flipping techniques to preprocess images to enhance features and improve the model's ability to recognize vehicles under different lighting conditions. Then, through the maximum pooling method and the introduction of attention mechanism, the system can extract key feature regions more efficiently. In addition, we use RPN (Region Proposal Networks) to select and classify target regions, and process the results through loss functions and cross-entropy loss functions. Finally, the SoftMax classifier is used to normalize the model output, and the probability prediction of each vehicle class is obtained. Compared to the existing Faster R-CNN, our system achieved an average accuracy of 94.7% on various vehicle identification, and the accuracy per sample was also improved. This system can effectively identify the categories of vehicles in the surveillance system, providing a new approach for researching object detection algorithms.