How YOLOv5 Powers Real‑Time City Management Video Analysis

Object detection is a core technique in computer vision, widely used in smart video surveillance, autonomous driving, and industrial inspection. In a city‑management video analysis project, a custom dataset and experience are leveraged to automatically recognize and report events, reducing labor costs.

Overall Recognition Process

The system captures video frames, runs a target detection model to generate bounding boxes for objects such as people, vehicles, and street vendors, and then applies logical rules to produce event images, e.g., an unlicensed street vendor incident.

The detection algorithm is the key component that transforms raw frames into structured event data.

Target Detection Algorithm Choices

Two‑stage detectors offer high accuracy but are too slow for real‑time city‑management needs.

Transformer‑based detectors struggle with small objects and also have insufficient speed.

One‑stage detectors sacrifice a small amount of accuracy for a large speed gain; the project therefore adopts a one‑stage model, specifically YOLOv5.

YOLOv5 Architecture

YOLOv5 consists of four main parts: input, backbone network, neck (FPN + PAN), and prediction output.

Input and Mosaic Augmentation

Raw images are first processed with the Mosaic augmentation method, which randomly crops four images and stitches them together. This enriches background diversity, improves model stability, and effectively increases the batch size.

Backbone Network

The backbone extracts features from the augmented image, using a lightweight architecture to keep inference speed high while preserving detection accuracy.

Neck (FPN + PAN)

The feature maps from the backbone are fed into a neck built with Feature Pyramid Network (FPN) and Path Aggregation Network (PAN). This enriches semantic information and improves localization, boosting overall detection performance.

Prediction Output

The neck output passes through a convolutional layer producing three scale tensors: bs*3*80*80*(numclass+5), bs*3*40*40*(numclass+5), and bs*3*20*20*(numclass+5). Bounding boxes are decoded from these tensors, then filtered with Non‑Maximum Suppression (NMS) to obtain the final detection results for city‑management events.

Detection Results

The processed frame shows detected objects with confidence scores. Logical rules then identify whether the frame contains an event, such as an unlicensed street vendor, and generate the corresponding event image.

Conclusion

The project demonstrates that YOLOv5 provides a balanced solution for city‑management video analytics, delivering both high detection accuracy and real‑time performance. While this overview covers the main components, deeper implementation details and the extensive research literature behind each part merit further study.