INTELLIGENT ROAD ACCIDENT DETECTION SYSTEM BASED ON HEAT MAPS AND NEURAL NETWORK OBJECT DETECTION
Authors:
1.
Kazan State Energy University
2.
Kazan National Research Technical University named after A.N. Tupolev
3.
KNITU-KAI
4.
Akademiya nauk Respubliki Tatarstan
GBU «Bezopasnost' dorozhnogo dvizheniya»
GBU «Bezopasnost' dorozhnogo dvizheniya»
5.
KNITU im. A.N. Tupoleva-KAI:
6.
Kazan National Research Technical University named after A.N. Tupolev
Type:
Article
Pages:
from 132
to 136
Status:
Published
Received:
10.02.2026
Accepted:
19.02.2026
Published:
26.02.2026
Language:
Russian
Keywords:
ROAD ACCIDENT, YOLO, HEAT MAP, COMPUTER VISION, VIDEO SURVEILLANCE, TRAFFIC ANOMALIES, TRAFFIC FLOW, NEURAL NETWORK, OBJECT DETECTION
Abstract:
This article describes a system for automatically detecting road accidents using heat maps and neural network object detection. Object detection in road scenes is typically implemented using the YOLO, SSD, and Faster R-CNN models. Heat maps are widely used in video surveillance to visualize the density and dynamics of object movement. The combined use of heat maps and neural network object detection improves the information content of road scene analysis and increases the proportion of correctly identified incidents. The developed system is based on a combination of object detection, frame activity analysis, and heat map detection. The system architecture includes the following functional blocks: a video stream source, an object detection module, a motion heat map construction module, anomaly analysis module, and a video storage module. Video stream sources can include both RTSP cameras and local video recordings, with the ability to process multiple streams in parallel. The object detection module for road scene analysis utilizes the YOLO11n convolutional neural network model, ensuring high performance with acceptable detection accuracy in real time. A heat map is generated using a two-dimensional matrix, each cell of which reflects the intensity of vehicle activity in the corresponding area of the frame. The anomaly analysis module identifies anomaly zones on the heat map. When signs of an accident are detected on the heat map, the video storage module is triggered. The system initiates recording of a video fragment up to 5 minutes long. Recording ends after a specified time has elapsed or when the anomaly disappears. The system is implemented in Python using computer vision and neural network analysis libraries. Experiments have shown that system performance depends on the video frame processing mode: using only a CPU, processing speeds of approximately 2-5 frames per second for a single video stream are achieved, while using a GPU ensures near-real-time processing. The system's approach enables the accurate identification of emergency situations involving prolonged traffic congestion, disruptions to natural traffic patterns, and vehicle stops in unusual areas. The proposed approach can be used in Smart City systems, transport monitoring centers, and automated dispatch control systems.
This article describes a system for automatically detecting road accidents using heat maps and neural network object detection. Object detection in road scenes is typically implemented using the YOLO, SSD, and Faster R-CNN models. Heat maps are widely used in video surveillance to visualize the density and dynamics of object movement. The combined use of heat maps and neural network object detection improves the information content of road scene analysis and increases the proportion of correctly identified incidents. The developed system is based on a combination of object detection, frame activity analysis, and heat map detection. The system architecture includes the following functional blocks: a video stream source, an object detection module, a motion heat map construction module, anomaly analysis module, and a video storage module. Video stream sources can include both RTSP cameras and local video recordings, with the ability to process multiple streams in parallel. The object detection module for road scene analysis utilizes the YOLO11n convolutional neural network model, ensuring high performance with acceptable detection accuracy in real time. A heat map is generated using a two-dimensional matrix, each cell of which reflects the intensity of vehicle activity in the corresponding area of the frame. The anomaly analysis module identifies anomaly zones on the heat map. When signs of an accident are detected on the heat map, the video storage module is triggered. The system initiates recording of a video fragment up to 5 minutes long. Recording ends after a specified time has elapsed or when the anomaly disappears. The system is implemented in Python using computer vision and neural network analysis libraries. Experiments have shown that system performance depends on the video frame processing mode: using only a CPU, processing speeds of approximately 2-5 frames per second for a single video stream are achieved, while using a GPU ensures near-real-time processing. The system's approach enables the accurate identification of emergency situations involving prolonged traffic congestion, disruptions to natural traffic patterns, and vehicle stops in unusual areas. The proposed approach can be used in Smart City systems, transport monitoring centers, and automated dispatch control systems.
Keywords:
ROAD ACCIDENT, YOLO, HEAT MAP, COMPUTER VISION, VIDEO SURVEILLANCE, TRAFFIC ANOMALIES, TRAFFIC FLOW, NEURAL NETWORK, OBJECT DETECTION
ROAD ACCIDENT, YOLO, HEAT MAP, COMPUTER VISION, VIDEO SURVEILLANCE, TRAFFIC ANOMALIES, TRAFFIC FLOW, NEURAL NETWORK, OBJECT DETECTION
