How to Build a Real-Time Deep Learning Object Detector with OpenCV and Python

Using Deep Learning and OpenCV for Video Object Detection

To build a real‑time deep‑learning object detector with OpenCV, we first need to connect a camera or video stream and apply detection to each frame.

Setup and Required Packages

Start by creating a file named real_time_object_detection.py and import the necessary packages (lines 2‑8). Ensure you have imutils and OpenCV 3.3 (or newer) with the contrib modules installed in your Python virtual environment.

Note: Install OpenCV 3.3+ and the opencv-contrib package to obtain the DNN module.

Command‑Line Arguments

The script accepts the following arguments (the image‑specific arguments from the original project are omitted):

--prototxt : path to the Caffe prototxt file.

--model : path to the pretrained model.

--confidence : minimum probability threshold to filter weak detections (default 0.2).

Initializing Classes and Colors

Lines 22‑26 initialize the CLASS labels and random COLORS used for drawing bounding boxes.

Loading the Model and Video Stream

Lines 30‑34 load the serialized model (prototxt and weights) and set up the video source. The VideoStream class (from imutils) is started, the camera is warmed up, and an FPS counter is created.

Processing Frames

For each frame (starting at line 43), the script resizes the frame, constructs a blob (line 48), and feeds it to the network (lines 52‑53) to obtain detections.

It then iterates over the detections, extracts the confidence (line 59), filters by the confidence threshold (line 63), obtains the class index (line 67) and bounding‑box coordinates (lines 68‑69). A label combining the class name and confidence is built (lines 72‑73) and drawn with a colored rectangle (lines 74‑75). The label is placed above the box unless there is insufficient space (line 76), and the text color is adjusted (lines 77‑78).

The loop also displays the frame (line 81), checks for the “q” key to quit (lines 82‑86), and updates the FPS counter (line 89).

Cleanup

When the loop exits, the FPS counter stops (line 92) and the final FPS is printed (lines 93‑94). The display window is closed (line 97) and the video stream is stopped (line 98).

Results

Running the script on a webcam yields a video stream with real‑time detections, typically achieving 6‑8 FPS depending on hardware. The detector can recognize people, sofas, chairs, and other objects.

Performance Tips

Skip frames to increase speed.

Use faster MobileNet variants (trading accuracy for speed).

Try quantized SqueezeNet models for a smaller footprint.