On-road vehicle and pedestrian detection using improved codebook model

In this paper, an improved implicit shape model is presented for on-road vehicle and pedestrian detection. Implicit shape model (ISM) is widely used for object detection and categorization. The training of ISM usually consists of three components: interest point detector, local feature descriptor, codebook generation. We evaluate six common interest point detectors to determine the best detector for vehicles and pedestrians, and the experiments show that Harris Detector is more efficient than the others. The original shape context local feature descriptor is sensitive to shape with points near boundaries of bins, as each point gives hard distribution to the bin. Therefore, a fuzzy function is employed to make each point gives soft distribution to all around bins to make it robust to shapes with small difference on boundaries of bins. Finally, k-means algorithm is replaced by Mean shift to generate codebook, as it produces more accurate codebook on datasets without small bandwidth.