Muhammet Sait Bozdemir

The system combines YOLOv8 detection with ByteTrack tracking so objects retain consistent identities across video frames. Users define a four-corner region of interest, allowing the application to distinguish pedestrians inside and outside the area and count vehicles without repeated detections.

It supports webcam, single-video, and batch processing, with configurable model settings, color-coded overlays, tracking IDs, vehicle counts, and FPS monitoring.

ReLU activations were replaced with calibrated LIF neurons, BatchNorm parameters were folded into convolution layers, and classification was performed through accumulated output spikes. Channel-wise thresholds helped account for different activation ranges across the network.

The ANN baseline reached 95.88% accuracy. The converted SNN reached 95.00% at 128 timesteps and 95.75% at 256 timesteps. The project also exports model artifacts for hardware-oriented implementation.

The project transforms hourly measurements into daily time-series features and implements the SVM dual optimization problem with a quadratic-programming solver. It compares linear, RBF, polynomial, and custom trend-aware kernels using cross-validation.

The work reinforced an important modeling lesson: accuracy alone was misleading for the imbalanced pollution dataset, so recall and F1 were necessary to compare models responsibly.

Three digit images are combined and paired with captions such as “one hundred twenty-three.” A pretrained ViT-Tiny image encoder and DistilBERT text encoder project both modalities into a shared embedding space trained with SigLIP contrastive loss.

The project covers multimodal dataset generation, frozen pretrained encoders, trainable projection heads, contrastive training, and Top-1/Top-5 retrieval evaluation.