TrackZone con Ultralytics YOLO26
¿Qué es TrackZone?#
TrackZone se especializa en supervisar objetos dentro de zonas concretas de un fotograma, en lugar de hacerlo en todo el fotograma. Basado en Ultralytics YOLO26, integra la detección de objetos y el seguimiento específicamente en zonas de vídeos y señales de cámaras en directo. Como YOLO26 ejecuta detección y seguimiento mediante aprendizaje profundo en tiempo real, TrackZone es adecuado para aplicaciones como la supervisión de multitudes y la vigilancia.
Ver: Cómo seguir objetos en una región con Ultralytics YOLO26 | TrackZone 🚀
Ventajas del seguimiento de objetos por zonas (TrackZone)#
- Análisis específico: El seguimiento de objetos en zonas concretas permite obtener información más precisa y supervisar y analizar con exactitud las áreas de interés, como los puntos de entrada o las zonas restringidas.
- Menos carga de trabajo posterior: Al ignorar los objetos que quedan fuera de la zona, TrackZone elimina las detecciones irrelevantes, por lo que hay menos objetos que contar, registrar o sobre los que generar alertas en la lógica que desarrolles a partir de estos resultados. La detección sigue ejecutándose en un fotograma enmascarado de tamaño completo, en lugar de recortar estrictamente la zona, así que la ventaja es obtener resultados más limpios y específicos, no acelerar la inferencia del modelo.
- Mayor seguridad: El seguimiento por zonas mejora la vigilancia al supervisar áreas críticas, lo que ayuda a detectar antes actividades inusuales o brechas de seguridad.
- Soluciones escalables: La capacidad de centrarse en zonas específicas hace que TrackZone se adapte a distintos escenarios, desde espacios comerciales hasta entornos industriales.
Aplicaciones en el mundo real#
| Agricultura | Transporte |
|---|---|
![]() | ![]() |
| Seguimiento de plantas en un campo con Ultralytics YOLO26 | Seguimiento de vehículos en carretera con Ultralytics YOLO26 |
# Run a trackzone example
yolo solutions trackzone show=True
# Pass a source video
yolo solutions trackzone source="path/to/video.mp4" show=True
# Pass region coordinates
yolo solutions trackzone show=True region="[(150, 150), (1130, 150), (1130, 570), (150, 570)]"TrackZone utiliza la lista region para saber qué parte del fotograma debe supervisar. Define el polígono para que coincida con la zona física que te interesa (puertas, portones, etc.) y mantén show=True activado durante la configuración para comprobar que la superposición se alinea con la señal de vídeo.
- Cada entrada de
regiones una coordenada de píxel(x, y)en el fotograma de vídeo. Enumera los puntos en el orden en que deben conectarse alrededor del perímetro de la zona que quieras supervisar. - Las coordenadas dependen de la resolución del fotograma, por lo que una región configurada para una señal de 1280×720 no se alineará con otra de 640×480. Mantén
show=Trueactivado durante la configuración para confirmar que la superposición coincide con tu señal. TrackZonereduce los puntos a su envolvente convexa, por lo que una forma cóncava se simplifica al polígono convexo más pequeño que contiene todos sus puntos. Para formas no convexas o varias zonas separadas, usa la solución RegionCounter.- Si omites por completo
region, se usa una zona predeterminada de[(75, 75), (565, 75), (565, 285), (75, 285)].
Argumentos de TrackZone#
Aquí tienes una tabla con los argumentos de TrackZone:
| Argumento | Tipo | Predeterminado | Descripción |
|---|---|---|---|
model | str | None | Ruta a un archivo de modelo Ultralytics YOLO. |
region | list o dict | None | Puntos que definen la región de interés, ya sea una lista de tuplas (x, y) o un diccionario que asigna nombres de región a listas de puntos para varias regiones (solo RegionCounter). Si None, las soluciones que requieren una región usan una región predeterminada predefinida. |
line_width | int | 2 | Grosor de línea de los cuadros, puntos clave y recuentos que dibuja la solución. |
verbose | bool | True | Activa el registro por fotograma de la forma de entrada, los recuentos por clase y la velocidad de procesamiento de la solución. La llamada de seguimiento en sí siempre es silenciosa. |
La solución TrackZone admite los parámetros track:
| Argumento | Tipo | Predeterminado | Descripción |
|---|---|---|---|
tracker | str | 'tracktrack.yaml' | Especifica el algoritmo de seguimiento que se va a usar. Opciones integradas: botsort.yaml, bytetrack.yaml, ocsort.yaml, deepocsort.yaml, fasttrack.yaml, tracktrack.yaml. |
conf | float | 0.25 | Establece el umbral de confianza de las detecciones; los valores más bajos permiten seguir más objetos, pero pueden incluir falsos positivos. |
iou | float | 0.5 | Establece el umbral de intersección sobre unión (IoU) para filtrar las detecciones solapadas. |
classes | list | None | Filtra los resultados por índice de clase. Por ejemplo, classes=[0, 2, 3] solo sigue las clases especificadas. |
device | str | None | Especifica el dispositivo para la inferencia (p. ej., cpu, cuda:0 o 0). Permite elegir entre CPU, una GPU concreta u otros dispositivos de cálculo para ejecutar el modelo. |
imgsz | int or tuple | 640 | Tamaño de imagen de destino para la inferencia: un entero para una imagen cuadrada o una pareja (height, width). Las dimensiones se ajustan al stride del modelo. |
max_det | int | 300 | Número máximo de detecciones que se conservan por fotograma antes del seguimiento, lo que limita los objetos que procesa una solución en escenas densas. |
quantize | int or str | None | Precisión de inferencia: 16 (FP16) o 32/unset (FP32) para los modelos PyTorch y TorchScript fuera de CPU; los demás formatos usan la precisión que seleccionen el artefacto y el entorno de ejecución. Sustituye al argumento obsoleto half. |
Además, están disponibles las siguientes opciones de visualización:
| Argumento | Tipo | Predeterminado | Descripción |
|---|---|---|---|
show | bool | False | Si True, muestra las imágenes o los vídeos anotados en una ventana. Es útil para obtener información visual inmediata durante el desarrollo o las pruebas. |
show_conf | bool | True | Muestra la puntuación de confianza de cada detección junto a su etiqueta. Permite conocer el grado de certeza del modelo para cada detección. |
show_labels | bool | True | Muestra etiquetas para cada detección en la salida visual. Facilita la identificación inmediata de los objetos detectados. |
Cuenta los objetos dentro de la zona#
Cada llamada al seguidor devuelve un objeto SolutionResults cuyo atributo total_tracks contiene el número de objetos que se están siguiendo actualmente dentro de la zona. Léelo en cada fotograma para supervisar la ocupación en directo, por ejemplo, para registrar el nivel de actividad de un punto de entrada o una zona restringida:
import cv2
from ultralytics import solutions
cap = cv2.VideoCapture("path/to/video.mp4")
assert cap.isOpened(), "Error reading video file"
region_points = [(150, 150), (1130, 150), (1130, 570), (150, 570)]
trackzone = solutions.TrackZone(show=False, region=region_points, model="yolo26n.pt")
while cap.isOpened():
success, im0 = cap.read()
if not success:
break
results = trackzone(im0)
print(f"Objects currently in zone: {results.total_tracks}") # live zone occupancy
cap.release()Preguntas frecuentes#
Ejecuta el siguiente comando para seguir objetos solo dentro de la zona y pasa
regionpara establecer las coordenadas de la zona. Esto es adecuado para aplicaciones como la vigilancia, la gestión de multitudes o cualquier escenario que requiera seguimiento por zonas.yolo solutions trackzone source="path/to/video.mp4" show=TrueCon solo unas pocas líneas de código, puedes configurar el seguimiento de objetos en zonas concretas e integrarlo fácilmente en tus proyectos.
import cv2 from ultralytics import solutions cap = cv2.VideoCapture("path/to/video.mp4") assert cap.isOpened(), "Error reading video file" w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS)) video_writer = cv2.VideoWriter("trackzone_output.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h)) trackzone = solutions.TrackZone( show=True, region=[(150, 150), (1130, 150), (1130, 570), (150, 570)], model="yolo26n.pt" ) while cap.isOpened(): success, im0 = cap.read() if not success: break results = trackzone(im0) video_writer.write(results.plot_im) cap.release() video_writer.release() cv2.destroyAllWindows()Configurar los puntos de la zona para procesar vídeo con Ultralytics TrackZone es sencillo y personalizable. Puedes definir y ajustar las zonas directamente mediante un script de Python, lo que te permite controlar con precisión las áreas que quieres supervisar.
from ultralytics import solutions # Define los puntos de la región region_points = [(150, 150), (1130, 150), (1130, 570), (150, 570)] # Inicializa TrackZone trackzone = solutions.TrackZone( show=True, # mostrar el resultado region=region_points, # pasar los puntos de la región )Recuerda que
TrackZonereduce los puntos a su envolvente convexa, por lo que un contorno cóncavo se simplifica al polígono convexo más pequeño que contiene todos sus puntos.Las tres soluciones funcionan con regiones, pero responden a preguntas distintas:
Solución Úsala para Resultado habitual TrackZone Seguir objetos y supervisar la ocupación en directo dentro de una única zona convexa ID de los objetos seguidos y total_tracksde la zonaObjectCounter Contar los objetos que cruzan una línea o entran y salen de una región Recuentos acumulados de entradas y salidas RegionCounter Contar objetos en una o varias regiones arbitrarias (incluidas las no convexas) Recuentos de objetos por región Elige TrackZone cuando quieras realizar un seguimiento continuo dentro de una zona y RegionCounter cuando necesites varias zonas o una forma no convexa.
Comentarios40
Haz una pregunta, comparte tus resultados o ayuda a otra persona.
Paul231200
Hello, great development, but I had a question about the advantages of your solution in comparison with the regions of interest that can be done in Opencv, also the question is whether it is planned to add this development to such models as yolov8?
1
Glenn JocherUltralytics(editado)@Paul231200 thank you for your question! TrackZone offers targeted and efficient region-based tracking, reducing computational overhead compared to general frame analysis in OpenCV. As for YOLOv8, TrackZone works with any Ultralytics model you pass via
model=, including YOLOv8, though we recommend YOLO26 for new projects. For more details, you can explore TrackZone here: https://docs.ultralytics.com/guides/trackzone/.
moshiour
Hi, Its a nice feature to yolo! How do make it counting on top of it? Thank you in advance!
1
Glenn JocherUltralytics(editado)@moshiour thank you for your kind words!
TrackZonereturns a results object withtotal_tracks, the number of objects currently tracked inside the zone:results = trackzone(im0)thenresults.total_tracks. For counting objects entering or leaving a region, see the Object Counting guide. Let us know if you have more questions!1RizwanMunawar
@Muhammad-237 Hi, We currently don't support the counting feature with TrackZone, but It's on our TODO's list, and we will release this feature soon! Thank you! In the meantime, you can explore our documentation available at: https://docs.ultralytics.com/guides/object-counting/#__tabbed_1_1
biancalepe
Thank you for this informative resource! I have a question - am I able to obtain the unique IDs of identified objects within 20 second windows? In this case, I'd like to know how many unique cars in total were in my zone for each 20 second time period of my video.
Below is my starting code:
def count_specific_classes(video_path, output_video_path, model_path, classes_to_count, region_points):
"""Count specific classes of objects in a video.""" cap = cv2.VideoCapture(video_path) assert cap.isOpened(), "Error reading video file" # Video writer w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS)) video_writer = cv2.VideoWriter(output_video_path, cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h)) # Initiate TrackZone (object tracking in zones, not complete frame) trackzone = solutions.TrackZone(show=True, # Display the output region=region_points, # pass the region points model=model_path, # use any model that Ultralytics supports classes=classes_to_count) # which objects to track # Process video while cap.isOpened(): success, im0 = cap.read() if not success: print("Video frame is empty or processing is complete.") break results = trackzone(im0) #print (results) # access the output video_writer.write(results.plot_im) # write the video output cap.release() video_writer.release() cv2.destroyAllWindows() # destroy all open windows
Glenn JocherUltralytics(editado)@biancalepe you can access unique track IDs from
trackzone.track_idsand implement time-based windowing for counting. Here's how to modify your code to count unique cars in 20-second intervals:def count_unique_cars_by_time_window(video_path, output_video_path, model_path, classes_to_count, region_points, window_seconds=20): cap = cv2.VideoCapture(video_path) assert cap.isOpened(), "Error reading video file" w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS)) video_writer = cv2.VideoWriter(output_video_path, cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h)) trackzone = solutions.TrackZone( show=True, region=region_points, model=model_path, classes=classes_to_count ) frame_count = 0 window_frame_count = window_seconds * fps # frames per 20-second window current_window = 1 unique_ids_in_window = set() window_results = [] while cap.isOpened(): success, im0 = cap.read() if not success: break results = trackzone(im0) # Track IDs of the objects currently inside the zone unique_ids_in_window.update(trackzone.track_ids) frame_count += 1 # Check if window is complete if frame_count >= window_frame_count: unique_count = len(unique_ids_in_window) window_results.append(f"Window {current_window} ({window_seconds - 20}-{window_seconds}s): {unique_count} unique cars") print(f"Window {current_window}: {unique_count} unique cars detected") # Reset for next window unique_ids_in_window.clear() current_window += 1 frame_count = 0 window_seconds += 20 video_writer.write(results.plot_im) # Handle final incomplete window if unique_ids_in_window: unique_count = len(unique_ids_in_window) print(f"Final window: {unique_count} unique cars detected") cap.release() video_writer.release() cv2.destroyAllWindows() return window_resultsThis tracks unique car IDs within each 20-second window and provides counts for analysis.
biancalepe
Thank you!
Running the code, I see that cars exist in the twenty second windows but the code returns back zero cars for all windows. As I was trying to parse out the code, I printed results.boxes and I get this error --> AttributeError: 'SolutionResults' object has no attribute 'boxes'. I get a similar error when I try to print out results.boxes.id. Printing results seems to get me a "total_tracks" and a "speed" option, but nothing else. Is there something I should be doing?
RizwanMunawar(editado)
@biancalepe If you want to retrieve the IDs of tracked objects in a specific Zone (these will be unique), you can use the code mentioned below. Please note SolutionResults class includes the data highlighted here. Track IDs can be extracted directly by Solution Object, i.e,
trackzone.track_ids. Thanksimport cv2 from ultralytics import solutions cap = cv2.VideoCapture("path/to/video.mp4") assert cap.isOpened(), "Error reading video file" # Define region points region_points = [(150, 150), (1130, 150), (1130, 570), (150, 570)] # Video writer w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS)) video_writer = cv2.VideoWriter("trackzone_output.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h)) # Init trackzone (object tracking in zones, not complete frame) trackzone = solutions.TrackZone( show=True, # display the output region=region_points, # pass region points model="yolo11n.pt", # use any model that Ultralytics support, i.e. YOLOv9, YOLOv10 # line_width=2, # adjust the line width for bounding boxes and text display ) # Process video while cap.isOpened(): success, im0 = cap.read() if not success: print("Video frame is empty or processing is complete.") break results = trackzone(im0) print(trackzone.track_ids) # access the track_ids video_writer.write(results.plot_im) # write the video file cap.release() video_writer.release() cv2.destroyAllWindows() # destroy all opened windowsbiancalepe
Sorry @muhammadrizwanmunawar the reason I was doing the print statements is because I was trying to figure out why the first block of code written by @glenn-jocher wasn't working. I isolated the problem to the extract IDs from results part. Using your code and glenn's I updated that section to be the code below, which seems to have fixed the problem:
Extract track IDs from results
if trackzone.track_ids is not None: track_ids = trackzone.track_ids unique_ids_in_window.update(track_ids)1
Glenn JocherUltralytics@biancalepe excellent troubleshooting! You've correctly identified that track IDs should be accessed via
trackzone.track_idsrather than from theresultsobject. This is the proper way to extract track IDs from TrackZone for time-windowed unique counting. Great job combining the solutions!
amoore1724
Can this be used to track two separate regions in the same frame?
RizwanMunawar(editado)
yes! Currently, we support the tracking objects in a single zone.
Caio PassosHi! I have a question. In the docs, it says TrackZone supports classes as parameters. I tried to limit TrackZone to only detect the class 0 (person).
heres the code, almost like the example in docs:
import cv2 from ultralytics import solutions, YOLO cap = cv2.VideoCapture( "/path_to_video.mp4" ) assert cap.isOpened(), "Error reading video file" # Define region points region_points = [(150, 150), (1130, 150), (1130, 570), (150, 570)] # Video writer w, h, fps = ( int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS) ) video_writer = cv2.VideoWriter( "trackzone_output.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h) ) # Init trackzone (object tracking in zones, not complete frame) trackzone = solutions.TrackZone( show=True, # display the output region=region_points, # pass region points model="yolo11n.pt", names=[0], # use any model that Ultralytics support, i.e. YOLOv9, YOLOv10 line_width=1, # adjust the line width for bounding boxes and text display ) print(f" model class names: {trackzone.model.names}") # Process video while cap.isOpened(): success, im0 = cap.read() if not success: print("Video frame is empty or processing is complete.") break results = trackzone(im0) # print(results) # access the output video_writer.write(results.plot_im) # write the video file cap.release() video_writer.release() cv2.destroyAllWindows() # destroy all opened windowsit runs without any issue, but still detecting other classes like backpacks, for example. Can you provide some extra details on how to limit detections for person class only?
RizwanMunawar
@caio-passos Hi, thanks for sharing it. The actual argument to set classes is
classes, notnames,as mentioned in the documentation here: https://docs.ultralytics.com/guides/trackzone/#trackzone-arguments. Let me know if you have any questions. Thanks
Caio PassosThanks!! I've changed it in the code:
trackzone = solutions.TrackZone( show=True, # display the output region=region_points, # pass region points model="yolo11n.pt", classes=[0], # use any model that Ultralytics support, i.e. YOLOv9, YOLOv10 line_width=1, # adjust the line width for bounding boxes and text display )But it led to this error, something related to the size of the mask or tensor:
Traceback (most recent call last): File "/home/tarzan/Documents/programing/hive/aeroporto/track.py", line 38, in <module> results = trackzone(im0) ^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/solutions/solutions.py", line 195, in __call__ result = self.process(*args, **kwargs) # Call the subclass-specific process method ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/solutions/trackzone.py", line 73, in process self.extract_tracks(masked_frame) File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/solutions/solutions.py", line 120, in extract_tracks self.tracks = self.model.track(source=im0, persist=True, classes=self.classes, **self.track_add_args) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/engine/model.py", line 593, in track return self.predict(source=source, stream=stream, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/engine/model.py", line 549, in predict return self.predictor.predict_cli(source=source) if is_cli else self.predictor(source=source, stream=stream) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/engine/predictor.py", line 218, in __call__ return list(self.stream_inference(source, model, *args, **kwargs)) # merge list of Result into one ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 36, in generator_context response = gen.send(None) ^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/engine/predictor.py", line 336, in stream_inference self.results = self.postprocess(preds, im, im0s) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/models/yolo/detect/predict.py", line 55, in postprocess preds = ops.non_max_suppression( ^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/tarzan/Documents/programing/hive/aeroporto/.venv/lib/python3.12/site-packages/ultralytics/utils/ops.py", line 304, in non_max_suppression x, xk = x[filt], xk[filt] ~^^^^^^ IndexError: The shape of the mask [64] at index 0 does not match the shape of the indexed tensor [62, 6] at index 0I had experienced something similar with QueueManager before, but I couldn't figure it out where or what went wrong
RizwanMunawar
@caio-passos, which ultralytics version are you using?
Caio Passosit's ultralytics version 8.3.114
RizwanMunawar
@caio-passos, please upgrade the package with
pip install -U ultralytics. This error is already fixed in the latest version.1
Caio PassosIt solved the issue, thank you!! @RizwanMunawar
1
Parthpatil0024
Is there a way to use botsort without using the track method. how to do it if i am using a ncnn model format.
skyceeee
import cv2 import gc from ultralytics import solutions
Kamerayı başlat
cap = cv2.VideoCapture(0) assert cap.isOpened(), "Kamera açılırken hata oluştu!"
Kamera çözünürlüğünü belirle
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640) cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
Takip bölgesini tanımla (görüntü çözünürlüğüne uygun olmalı)
region_points = [(50, 50), (590, 50), (590, 430), (50, 430)]
TrackZone nesnesini başlat
trackzone = solutions.TrackZone( show=True, region=region_points, model="C:/yazilim/staj/trackzone/yolov8n480.pt", # Daha küçük model kullanıldı imgsz=480, # Model giriş boyutu sınırlandırıldı )
Canlı görüntü işleme döngüsü
while cap.isOpened(): success, frame = cap.read() if not success: print("Kamera görüntüsü alınamadı!") break
# Görüntüyü yeniden boyutlandır frame = cv2.resize(frame, (640, 480)) # TrackZone ile işlem yap try: processed_frame = trackzone.trackzone(frame) except Exception as e: print(f"Takip sırasında bir hata oluştu: {e}") break # Görüntüyü göster cv2.imshow("TrackZone Canlı Takip", processed_frame) # Bellek temizliği yap gc.collect() # 'q' tuşuna basarak çıkış yapabilirsiniz if cv2.waitKey(10) & 0xFF == ord('q'): breakKamera ve pencereyi serbest bırak
cap.release() cv2.destroyAllWindows()
This is my code. I use my own trained yolov8n model. When the camera is turned on, the area I drew also appears. No problems until then. But the moment a person enters the rectangle, the kernel dies. What could be the reason for this?My model does human detection. Error: The Kernel crashed while executing code in the current cell or a previous cell. Please review the code in the cell(s) to identify a possible cause of the failure. Click here for more info. View Jupyter log for further details.
Glenn JocherUltralyticsThe kernel crash could be due to insufficient memory or a compatibility issue with your custom YOLOv8 model. Ensure your system has enough resources available, and verify that the model file is correctly configured for the
TrackZonesolution. Additionally, try reducing the input resolution further (e.g.,imgsz=320) to minimize memory usage. For further debugging, check the full error log and confirm that your model works with standardpredictfunctionality outside ofTrackZone.
nichamond143
Hi, I'm using TrackZone with a custom model, but it's not detecting the target object. However, when I crop the zone of interest from the example video and put it through the model it works fine. I also trained the model on zoomed in images of the object because the full-frame image is too noisey and the object is very small. Any advice on how to approach this?
Glenn JocherUltralytics@nichamond143 it sounds like the model's training on zoomed-in images may limit its ability to generalize to full-frame inputs. You could try pre-processing the video frames to crop and resize the zone of interest before passing them to the model. Alternatively, consider retraining the model with multi-scale images, including both zoomed-in and contextual full-frame examples, to improve its detection robustness.
UltralyticsAssistantUltralytics
👋 Hello, thank you for your interest in Ultralytics 🚀! It sounds like you're exploring how TrackZone uses YOLO models for object tracking and analysis – fascinating topic!
If this is a ❓ Question about custom implementations, feel free to share more details about your use case so our team or the community can better assist. If it's a 🐛 Bug Report or you need specific technical help, please provide a clear description, along with any code snippets, error messages, or logs. Including a minimum reproducible example (MRE) would greatly help us debug and offer guidance swiftly.
For general inquiries or tips on achieving optimal results, consider reviewing our resources on model training and usage. Additionally, you can join the Ultralytics community on platforms like Discord, Reddit, or Discourse for deeper interactions and advice.
Lastly, please ensure you're running the latest version of the
ultralyticspackage and all required dependencies to avoid encountering previously resolved issues. You can upgrade using thepip install -U ultralyticscommand.Thank you for being part of the Ultralytics community! 😊 An Ultralytics engineer will be here to assist you further shortly.
skyceeee
By converting the yolo model to tensorrt and setting coordinates in Jetson Nano, I track objects in a certain area with Jetson Nano. Can I use trackzone instead of adjusting coordinates myself? Is Jetson Nano supported?
Glenn JocherUltralytics@skyceeee yes, you can use TrackZone with Jetson Nano for targeted object tracking within specific areas, eliminating the need to adjust coordinates manually. TrackZone is compatible with YOLO models, including TensorRT-exported ones. Ensure you optimize your model for Jetson Nano's limited resources for best performance. More details on TrackZone can be found here.
1
skyceeee
Thanks so much for this feature. Is it possible to use this feature with tensorrt format in Jetson nano?
Glenn JocherUltralytics@skyceeee thank you for your question! Yes, you can use the TensorRT-exported YOLO models on the Jetson Nano, including for the TrackZone feature. First, export your YOLO11 model to TensorRT format using the
yolo exportcommand. Then, ensure your TensorRT runtime is properly set up on the Nano, and integrate the exported model into your tracking workflow. For more details, refer to the Export Documentation. Let us know if you encounter any issues!1

